Assumptions in Structural Equation Modeling Rex B. Kline uctid aught ne that witout assumptions there i no proof. Therefore, in any argument examine the assumptions “E.T BELL (183-1960), Scottish mathematician and author ‘The least questioned assumptions are often the most questionable. "= Pleexe Patt Broca (1824-1880), French physician and anatomist ‘Teco taconite doce eenan te importance of assumptions in science. The main pur ‘poses of this chapter are to (I) elucidate the assump- ‘ions that underlie the application of structural equation modeling (SEM) and (2) emphasize the critical role of assumptions in specification, analysis, and interpreta- tion. The main goal ofthis presentation isto help read- ers appreciate just how much the substantive value of basically all phases of SEM depend on the veracity Of the underlying assumptions. T also argue that as sumptions are given short shrift in perhaps most SEM studies. This inattention implies thatthe interpretation ‘of results in the typical SEM analysis may be unwar- ranted, to. Many researchers in the behavioral sciences use standard statistical techniques that make certain as- sumptions about the data or model. For example, statis- ‘ical tests in multiple regression (MR) generally assume that the residuals are normally distributed and have ‘uniform variances across all levels of the predictors (homoscedasticty). Although it is possible to model curvilinear relations or interactive effects in multiple regression by entering the appropriate power or product term into the equation (e.g. X? for a quadratic effect ‘when predictor Xs also inthe equation; Cohen, Cohen, West, & Aiken, 2003), a standard regression analysis assumes linear relations only. The analysis of variance (ANOVA) generally requires normal population distri- butions, homogeneity of variance, and both equal and ‘uncorrelated error variances when analyzing scores from repeated-measures factors. 1 would wager that ‘most researchers could name the majority of the MR or ANOVA assumptions just listed. Unfortunately, there is evidence that too niany researchers fail to practice ‘what methodologists preach: that i, these assumptions are not taken seriously. For example, few authors of studies in which ANOVA is used provide any evidence about whether distributional or other assumptions were ‘met (Keselman et al, 1998). Without verification of assumptions, the results of statistical tsts—and cor- responding interpretations based on them—should be viewed with caution, Although the F-test in ANOVA is relatively robust against violation of normality or ho- ‘mogeneity assumptions, this is generally tue for large, representative samples with equal group sizes. Other- ‘wise, even relatively minor departures from normality ‘or homogeneity can seriously bias outcomes of the F= test (and the rest, too) (Wilcox, 1998). aan1128 ‘Compared with standard statistical techniques, there fare even more assumptions in a typical application of, SEM. These assumptions concern the model, infer- fences about the directionality of effects in structural ‘models or measurement models, and the data, The latter ‘concern distributional and other assumptions required by a particular estimation method, of which the most widely used (and default) method is maximurm likei- hood (ML) estimation. Each set of assumptions just mentioned is considered next. Major take-home mes- sages of this discussion are that (I) researchers must be even more vigilant about assumptions in SEM com- pared with using more standard statistical techniques; however, (2) if researchers are just as lackadaisical bout verifying assumptions in SEM as they seem to bbe when using simpler techniques like ANOVA or MR, then results from SEM studies may have litle interpre. tative value, DIRECTIONALITY ASSUMPTIONS ‘The specification of directionalites of presumed ca- sual effects, or “effect priority,” is represented in most structural equation models described in the literature. Directionality specifications concern both structural models and measurement models. Briefly, “structural ‘models” in SEM represent hypotheses about presumed direct or indirect causal effects among variables mea- sured either simultaneously (cross-sectional designs) or at different points in time (longitudinal designs). For ‘example, the hypotheses that verbal skills affect delin- ‘quency both directly and indirectly through thei prior impact on scholastic achievement would be represented inthe diagram of a structural model with the paths Verbal Skills ~> Delinquency and ‘Verbal Skills > Achievement -> Delinquency Where the presumed direct effects are represented by the Tine with a single arrowhead. The original SEM technique of path analysis, which dates (0 the 1920s and the work of geneticist Sewall Wright, concerns the ‘analysis of structural models where each observed vari- able isthe single indicator (measure) of an underlying Construct, or latent variable. It is also possible in SEM {o specify structural regression models—also known as “latent variable path models’—in which some con- structs are measured with multiple indicators (Kline, 2010, Chap. 5). Confirmatory factor analysis (CFA) II, FUNDAMENTALS models also feature the measurement of latent vati- ables, each with multiple indicators. ‘A “measurement model” is defined by (1) the dis- tinction between indicators and their corresponding constructs, and (2) specification of directional effects between observed and latent variables. In a standard ‘measurement model, scores on observed variables are presumed to be caused by hypothetical constructs, For example, if a vocabulary testis specified as an indica- tor of a latent verbal skills factor, then the specifica- tion ‘Verbal Skills > Vocabulary in the diagram of a measurement model reflects the hypothesis that vocabulary test scores reflect the in- fluence of an underlying verbal skills construct. The presumption that constructs affect observed scores, and not vice versa, corresponds toa traditional view of ‘measurement, also known as “reflective measurement.” Classical measurement theory assumes reflective mea: surement. For example, the fundamental equation of traditional reliability theory is XaTsE cy where X represents an observed score, T stands for a true score on the construct, and £ signifies a random error component that is normally distributed with a ‘mean of zero, Most measurement models analyzed in ‘SEM studies—and all models tested with CEA—as- ‘sume reflective measurement, I is not always appropriate to assume that indica- tors are caused by underlying factors. Consider this ‘example by Bollen and Lennox (1991): The variables ‘income, education, and oecupation are used to measure socioeconomic status (SES). In a standard measure- ‘ment model, these observed variables would be speci- fied as caused by a single underlying SES factor, or SES — Income, SES —> Education, and SES > Occupation ‘But we usually think of SES as the outcome of these variables (and probably others), not vice versa, For ex- ample, a change in any one of these indicators, such as salary increase, may affect SES. From the perspective of “formative measurement,” SES is a composite that is determined by its indicators, not the other way around that is,7. Assumptions in SEN Income -> SES, Education — SES, ‘and Occupation —» SES ‘The assumption of formative measurement reverses the presumed directionality of effects between indica- tors and latent variables relative to that presumed in reflective measurement models, and vice versa. Other implications of specifying reflective versus formative measurement are considered later, Discussed next are directionality assumptions of structural models in SEM. Assumptions of Structural Models Five general conditions must be met before one can rea- sonably infer a causal relation between two variables: 1. The presumed cause (e.g, X) must occur before the presumed effect (e.g. 1) that is, there is tem- poral precedence, 2. There is association, or an observed covaristion, between X and ¥. 3. There is isolation, which means that there are no other plausible explanations (eg, extraneous or confounding variables) of the covariation be- toveen X and ¥; that is, their statistical association hholds controlling for other variables that may also effect ¥ 4, The form ofthe distribution ofthe data is known; thats, the observed distributions match those as- sumed by the method used to estimate associa- tions. ‘5. The direction of the causal relation is correctly specified; thats, X indeed causes Yinstead of the reverse, or X and Y eause each other in a recipro- cal manner, ‘The requirement for temporal precedence is ad: dressed in experimental studies through random as signment of cases to conditions and in nonexperimen- tal studies through the use of longitudinal designs, also known as “panel designs,” where the putative cause and effect are each measured at different points over time. If all variables are measured simultaneously —which is rue in most SEM studies—then ic is not possible to demonstrate temporal precedence. In this case, a clear, substantive rationale is needed for specifying that X in fact is a cause of ¥. This rationale should have a solid 2113 basis in both theory and results of empitical studies. I should also identify other relevant variables to measure ‘and control for in statistical analysis (Second and third requirements). Without such a rationale, any inference about eausation in @ nonexperimental design has little justification. Data-related assumptions (fourth require- ‘ment) are considered later, but a gross mismatch be- tween data characteristics and statistical assumptions ‘of a particular method jeopardize inference accuracy. ‘The fifth requirement for causal inference may be the toughest, and it highlights an aspect of the logic in causal modeling that is circular: Interpretation of a large coefficient fr the path X —> Y as evidence for cau sation assumes that the true direction of causality actu ally “ows” from X to ¥. That is, we must assume thet the specification X > Ys correct in the frst place (and also thatthe other four requirements listed earlier are met). This isa critical characteristic of SEM, one not fully appreciated by all researchers: Interpretation of Statistical estimates of direct effects as indicating cau- sality assumes that the researcher's model is correct. This isa huge assumption, one that should take our col- lective breath away and give us pause before interpret- ing results of SEM analyses as evidence for causation. Tis necessary to assume tha the model is correct he- ‘cause the statistical association between X and ¥ could be strong even if (1) the true direct effect isin the oppo- site direction (i. ¥ causes X), (2) the two variables are ‘causes and effects of each other in a feedback loop, or G) there is no causal relation, but X and ¥ have a com- ‘mon cause, which gives rise to a spurious association, ‘Thatis, the hypotheses of XOY YOK KEY and that of @ spurious (noncausal)sssociation could all be consistent with the same dato. Without presuming that X causes Viste only viable model, theres no basis for interpreting a large coeficient for the path X—> Y 1s evidence for this directional effec. This is why Pearl (2000) reminded us that “causal assumptions are pre= requisite for validating any causal conclusion” (p. 136) This insight echoes that of Wright (1923) from nearly 80 years earlier, when he noted that “prior knowledge ofthe causal relations is assumed as prerequisite inthe theory of path coefficients” (p. 240), tis also critical to know that Wright invented path analysis in order to estimate the magnitudes of effects ‘when the basic causal pathways were already known (ea, genetics). That is, given a true causal model,148 the technique of path analysis could be applied to es- timate it among observed variables. However, this is ‘not how we generally use path analysis or related SEM techniques for analyzing latent variables today. In the bbchavioral sciences, we rarely know the true causal model. Instead, we usually ypothesize a causal mode! and then we test that model using sample data, This context of use is very different from that of Wright. ‘Specifically, when the true causal model is unknown but our hypothesized mode! fits the data, about all we ‘can say is that our model is consistent with the data but we cannot claim that our model is proven. In this way, ‘SEM can be seen a disconfirmatory technique, one that cean help us to reject false models (those with poor fit to the data), but it basically never confirms a particular ‘model when the true model is unknown. In this sense, referring to SEM as “confirmatory” is misleading, es- pecially in the behavioral sciences where true causal pathways are rarely known. Exogeneity and Endogensity PPethaps the most basic specification in structural mod- els is the distinction between exogenous variables and endogenous variables, To summarize, the presumed causes of exogenous variables are not represented in structural models, Instead, (1) the causes of exogenous variables are unmeasured (or unknown as far as the model is concerned), and (2) exogenous variables are ‘considered free both to vary and to covary. The later (covariance, unanalyzed association) is often repre- sented in model diagrams by the symbol with a curved line with two arrowheads (4_)) that connects every pair of exogenous variables. In contrast, the presumed measured causes of endogenous variables are explicitly represented in the model, which also implies that en- ddogenous variables are not free to vary or covaty. Endogenous variables in diagrams have direct ef fects pointing to them from exogenous variables or ‘ther endogenous variables. The specification of at least one endogenous variable as a direct cause of an- ‘ther endogenous variable implies an indirect effect. If X is exogenous and ¥, and ¥, are endogenous, for stance, then the specification X—> ¥, —>Y, describes an indirect effect, where ¥, is presumed to mediate atleast part of the effect of X on ¥,, The statistical estimate of an indirect effect isthe product ofthe path coefficients forall the direct effects that make up the indirect effect. Kenny (2008) reminded us of the requirements listed Il, FUNDAMENTALS next for interpreting the product of path coefficients as ‘evidence for an indirect causal effect: |. A mediational model is a causal model, For ex: ‘ample, the specification X > ¥, —> ¥, assumes that variable ¥ is actually a cause of ¥, and not vice versa, It also assumes variable X is causally prior to both ¥, and Y, Iether assumption is in- correct, then the results of a mediational analysis are of litle value. 2. Mediation is not statistically defined. Instead, statistics such as products of direct effects can be used to evaluate a presumed mediational mode!. Kaplan (2004) notes that the distinction between ‘exogenous and endogenous variables is not solely con- cceptual. That is, just because a researcher claims that 4 Variable is exogenous does not make it so. Instead, ‘there are statistical consequences for claiming thatthe ‘exogenous variables do not depend on (are not caused by) the endogenous variables. These consequences are implied by the assumption of exogeneity, a concept that is familiar in the economics literature but is less well known in the behavioral sciences. Briefly, “exogeneity” means thatthe parameters of the conditional distribu- tions of the endogenous vatiables given the exogenous variables are unrelated to those that describe the distri- butions of the exogenous variables by themselves (-., their means and variances). That is, knowing a param- cer inthe marginal distributions ofthe exogenous vari- ables gives no information about values of parameters in the structural equation model (Kaplan, 2004). Ways to check whether weaker or stronger forms of exogene- xy may hold ina particular sample are considered later in the section “Data-Related Assumptions." Exogeneity also implies thatthe disturbances (error or residual variances) of the endogenous variables in 2 structural model are unrelated to the exogenous vari- ables, In other words, all omitted (unmeasured) causes of the endogenous variables are uncorrelated with the ‘whole set of exogenous variables. This assumption is re- ferred to as “pscudo-isolation” (Bollen, 1989) or “self Games, Mulaik, & Brett, 1982), and it the estimation of direct effects and disturbance variances holding omitted causes constant, There is & similar requirement in MR: the residuals in ordinary least squares estimation are calculated to be indepen: dent of the predictors, which implies that all omitted predictors are assumed to be independent of predictors7, Assumations in SEM in the equation. In SEM, exogeneity thus implies that ‘not only are directionality specifications correct, but also that no unmeasured cause has anything to do with the exogenous variables. These strong requirements emphasize the importance of correct specification of ‘the model inthe first place. Given the specification X —> ¥, the situation where 4 putative exogenous variable X actually covaties with the error term for Y is known as “endogeneity.” which indicates that (1 exogeneity does not hold and (2) va able X is not really exogenous. Misspecification of the rection of the causal relation between X and Y can lead to endogeneity. If ¥ actually causes X (Y > 2), then X is clearly not exogenous; a reciprocal relation between X and ¥ (X= ¥) implies the same thing, Here is an example by Antonakis, Bendahan, Jacquart, and Lalive (2010): Suppose that more police are hired in order to reduce the level of crime, However, if an it crease in crime leads to the decision to hire more po- lice, then the latter is not exogenous because the two variables reciprocally affect each other. Endogeneity can also arise when there is measurement error in an exogenous variable and that error isnot explicitly rep- resented in the model, a point that is elaborated later. ‘See Antonakis and colleagues (2010) for description of other sources of endogeneity Disturbances In most structural models tested in the behavioral sei ences, disturbances of the endogenous variables are assumed to be uncorrelated. Independence of error terms is signaled in model diagrams by the absence of the symbol fora covariance (t_}) that connects any pair of disturbances, This specification implies that no two endogenous variables share a common omitted cause: that is, all omitted causes are unrelated to each fther. Tt also implies that the observed correlations ‘among the endogenous variables can be explained by other measured variables in the model, exogenous or endogenous. This presumption is known as the “local independence assumption” that the endogenous var fables are independent, given che (correctly specified) Structural model. The assumption of no common omit: ted causes is both restrictive and probably unrealistic, especially for the types of outcome variables studied in the behavioral sciences. For example, it seems un- likely that omitted causes of reading outcomes would be unrelated to omitted causes of arithmetic outcomes among clementary schoolchildren, Also, the error a5 variances of repeated-measures variables may over. lap, which describes autocorrelation among repeated ‘measures variables. The specification of a disturbance covariance (for unstandardized variables) or a disturbance correlation (for standardized variables), suchas D, _} D>, reflects the assumption that the corresponding’ endogenous variables (Y; and ¥,) share at least one omitted cause Unlike unanalyzed associations between measured ex- ‘ogenous variables (eg. X; 9 X,), the inclusion of dis- turbance covariances in the model is not routine, This is so in past because the addition of each disturbance covariance to the mode! “costs” one degree of freedom (4f), and thus makes the mode! more complicated. If there are substantive reasons for specifying disturbance covariances, however, then it is better to estimate the ‘model with these terms than without them. This is because the constraint that a disturbance covariance is zera when there are common causes tends to redis- tribute this association toward the exogenous end ofthe model, which can result in biased estimates of direct effects. In structural models with latent variables, the omission of theoretically justifiable correlated residuals. ‘may notin some cases harm mode! fit, but their omis> sion could change the meaning of latent variables and thus lead to inaccurate results (Cole, Ciesla, & Steiger, 2007), Recursive and Nonrecursive Models “There are two types of structural models, recursive and nonrecursive. Both assume thatthe exogenous variables are unrelated to the disturbances of the endogenous variables (n0 endogeneity). They differ concering assumptions about direct effects among endogenous variables and whether endogenous variables specified 4s causes are allowed to covary with the disturbances of other endogenous variables specified as outcomes. In recursive models, all causal effects are unidirectional: tha is, there are no feedback loops, so none of the en- ddogenoas variables are specified as both causes and ef- fects ofeach other Recursive models may also have op- tional correlated disturbances, but only between pars of endogenous variables without direct effects berween them. "Nonrecursive models have feedback loops or they ‘may have optional correlated errors between pats of endogenous variables with direct effects between them. ‘A diect feedback loop involves just two endogenous ‘variables specified as both causes and effets of each116 = other (¢2, ¥; © ¥,). Indirect feedback loops involve three or more variables, such as YoOnoKoy "Note that variables involved in feedback loops are each actually each measured only once and also simultane- ‘ously, That is they are estimated with data from cross- sectional designs, not longitudinal (panel) designs. Dis- turbances of variables involved in feedback loops are often specified as correlated. This specification makes sense because if variables are presumed to mutually ‘cause each other, then it seems plausible to expect that they may have shared omitted causes. In fact, the presence of disturbance correlations in particular pat- terns in nonrecursive models with feedback loops helps to determine their identification status (Kline, 2010, Chapter 6). Presented in Figure 7.1 are two structural models.! ‘These models are path models because there is a single indicator of every construct, but the same principles apply when some constructs are measured with mul- tiple indicators and the whole model is a structural re- gression model. Observed variables in Figure 7. are represented with squares and the disturbances of endog- enous variables are represented with circles, Constants that scale the disturbances are represented in the figure as I's, The symbol in Figure 7.1 with the two-headed ceurved arrow that exits and reenters the same variable (77) represents the variance of an exogenous variable is free to vary). This symbol also appears next to the disturbances because the computer must estimate the error variance, which is a model parameter. The II, FUNDAMENTALS model of Figure 71(a) is recursive because no endog- ‘enous variable is represented as both a cause and effect ‘of another, directly or indirectly. There is a disturbance correlation in Figure 7.1(), but there is no direet ef feet between ¥, and Ys. The specification D, _t D, assumes that there is at least one common unmeasured cause of Y; and Ys ‘The mode in Figure 71(b) has both a feedback loop and a disturbance correlation between a pair of vari ables with direct effects between them (Y, =2 ¥,). The latter specitication (D, ¢_J D,) implies that a predictor ‘of Y, oF ¥, covaries with the disturbance of Y, This rmodelimplied association is represented in Figure 10) by the path DADDY, ‘The other predictor of Y isthe exogenous variable X>, but exogenous variables and disturbances are assumed tw be orthogonal. The model in Figure 7.1(b) also im. plies that a predictor of Y;, or ¥,covaries with the dis. turbance of ¥, or DADDY ‘These mode!-implied correlations between endogenous predictors and the disturbances of other endogenous ‘outcome variables violate the ordinary least squares sumption that the residuals are independent of all the predictors. This explains why standard MR is generally inappropriate for estimating direct effects in nonre- ‘cursive path models, but it i no particular problem for ‘ML estimation or special forms of regression, such as (@) Recursive (©) Noorecursive 2, 2 xt x vi 2 2 % % x» -——1% FIGURE 74. Examples of recursive and nonrecursive structural models7. Assumptions in SEM ‘two-stage least squares, that correct for model-implied ‘correlations between endogenous predictors and distur- ‘bances (Kline, 2010, pp. 156-157). ‘There are two special assumptions for nonrecursive structural models with feedback loops. Kaplan, Harik, and Hotchkiss (2001) remind us that data from a eross- sectional design give only a snapshot of an ongoing ddynamie process, Therefore, the estimation of recipro- cal effects in a feedback loop with cross-sectional data requires the assumption of “equilibrium.” This means that any changes in the system underlying a presumed feedback relation have already manifested their effects, ‘and that the system is in a steady state. That is, esti ‘mates of the direct effects that make up the feedback Joop do not depend on the particular time point of data collection. A second assumption is that of “stationari- «ty the requirement thatthe basie causal structure does ‘not change over time. Both assumptions just described fare demanding, and thus probably unrealistic. Also, there is no direct way to verify these assumptions inthe data. Instead, these assumptions should be evaluated on ‘ational grounds. However, the assumptions of equil rium and stationarity are rarely acknowledged in stud- ies where feedback effects are estimated with cross- sectional data. This is unfortunate because results of ‘computer simulation studies by Kaplan and colleagues indicate that violation of the equilibrium assumption ‘can Tead (0 severely biased estimates of the direct ef fects in feedback loops. Justitying Specifications in Structural Models ‘The specification of structural models in SEM relies heavily on the researcher's good judgment to (1) de- scribe and measure variables of interest, while avoid- ing the omission of causes correlated with those repre: sented inthe model; (2) correctly partition the variables into the mutually exclusive subsets of exogenous vari- ables and endogenous variables; (3) accurately lay out patterns of presumed direct and indiect effects; and (4) properly specify the error covariance structure (ic, dependent or correlated disturbances). This process is ‘based on multiple assumptions, the most overarching of ‘which i that the researcher's model is basically correct before any interpretation about causality ean be made. It also requires strong knowledge about the phenomena under study. The need for wisdom about the research problem is even greater when there is no temporal pre: cedence (all variables are measured simultaneously). Unfortunately, too many authors of SEM studies in a7 ‘which structural models are analyzed do not give tea- soned accounts oftheir specifications about exogeneity, directionality, and the structure ofthe ero erm. This is often apparent when the results section of an SEM utile is riuch longer then the introduction, ia which the rationale for model specification should be fllyex- plained. That is, too much attention is typically paid to the statistical machinations of SEM and not enough to the assumptions on which itis based. Examples of recent works where the authors presented quit detailed rationales for their directionality specifications (ce they are good role models) incade Sava (2002), who tesed a path model of student educational cutcomes, and Houghton and Jinkerson (2007), who tested a Stcuctural regression model of job satisfaction ‘Another serious shortcoming of SEM studies in which structural models are analyzed i the failure to consider equivalent models. Briefly, an “equivalent ‘model is based on the same variables asthe research- c's original model, has the same number of parame- ters, andi fitted to Same data. Both the original model and the equivalent version have the exact same fit the same data. Equivalent structural models are generated by applying the Lee-Hershberger replacing rules (2. Hershberger, 1994; Kline, 2010, pp. 225-228) such thatthe dicectonalites of certain paths are changed without affecting, model fit. Suppose that a structural rmodel inches the path ¥,—> ¥, and that both variables te specified to have common causes. In this cas, all ofthe following paths can be substituted for Y, —> Y5 ‘without changing the overall fit of the model: ¥5 — ¥, Dy \Y Dy and the equlity-constrained reciprocal e feet ¥; 2 Y, ity Y; > Yo=Y_~> ¥,). The respecited equivalent models say very different things about the directionality of effects between Y, and ¥, (if ay), but thy all have identical itt the data, Relatively simple strtural models may have few equivalent version, but more complicated ones may have hundreds or even thousands. However, most researchers fal to even ac- knowledge equivalent versions of their preferred struc- tural models, which is a form of confirmation bias ‘where alterative explanations ofthe same data are not recognized (Shah & Goldstein, 2006). This isa seri- cus problem, one that threatens the validity of perhaps ‘most SEM studies. There may also be nearequivalent ‘models that fit the same data just about a well asthe researcher's preferred model, but not exactly so. Near- equivalent models may be just as critical a validity threat as equivalent models, but near-equivalent models are rarely acknowledged in SEM studies.118 = Given these shortcomings, itis no wonder that many behavioral scientists are skeptical of the usefulness of ‘SEM for causal inference in nonexperimental designs (eg, Freedman, 191). Pearl (2009) attributes part of this “crisis” in SEM to the absence of a precise, math- ‘ematical language for formally evaluating causal hy- ppotheses embedded in (implied by) structural models. Pearl (Chapter, this volume) describes this framework {in more detail, soit is only briefly summarized here: It features the generation of implications of claims about causation represented in a structural model, and these claims are subjected to logical proofs that are indepen- dent of the data, The testable portions ofthese claims— that is, those derived from the parts of the model for ‘which there is no equivalent version—are represented bby graphs with parent and child nodes ofthe type ana- lyzed in graph theory, such as directed acyclic graphs. II, FUNDAMENTALS ‘These logical implications are then combined with the data in order to estimate Bayesian-type conditional probabilities of causal claims given the data and the original model. Assessment of global fit of the model to the data—which is tested with the model chi-square staistic—is deemphasized in Pear's (2009) framework in favor of local fitness testing of restrictions implied by testable causal hypotheses. This framework is not easy to learn and still depends on assumptions about cau- salty obtained from theoretical or empirical sources, but it offers greater rigor in the evaluation of testable causal hypotheses, including those about indirect ef- fects (mediational hypotheses). Time will tell whether Pearl (Chapter 5, this volume) is describing a next step in the evolution of SEM practice, but this framework ‘adds needed formalization and rigor to reasoning about causality in SEM. (a) Reflective (Latent — Indicator) (©) Formative (Indicator + Latent) x x x ? Latent (compost Yo. (¢) Formative (Indicator + Composite) x % % Composte, FIGURE 7.2. Examples of reflective and formative measurement models.7. Assumptions in SEM Assumptions of Measurement Models Considered next are ditectionality and other assump: tions of reflective measurement models and of forma- tive measurement models. Reflective Models Recall that reflective measurement models are the kind ‘most often analyzed in SEM, Presented in Figure 7.2(a) isa reflective model fora single factor with three dicators, X;-Xs, The constants in the igure (1's) scale the measurement errors and also the factor. Each ob- served variable in Figure 72a) is specified as an effect Indicator caused by an underlying latent variable that corresponds to a hypothetical construct and also by & measurement error term (e.g. for X,) that represents ‘unique variance. Thus, in reflective models (1) the dicators are specified as endogenous, and (2) measure- ‘ment error is represented atthe indicator level. Because the indicators are endogenous, their observed variances land covariances can be compared to values predicted by a reflective measurement model. It is generally as- ‘sumed thatthe factors and measurement errors in re- fective models are uncorrelsted; that is, any omitted ‘systematic cause of scores on the indicators has nothing 10 do with the factors: The factors themselves are as- ‘sumed to be continuous variables that represent a single ‘domain (ie. they are unidimensional) and are normally distributed. There are other statistical techniques, such as latent class analysis, that estimate categorical latent variables with levels that refer to membership in differ- ent inferred subpopulations, or classes, but SEM ana- Iyzes continuous latent variables only. independent esror terms are specified, as in Figure 7.2(@), then itis also assumed that omitted causes of different indicators are all pairwise uncorrelated. This {is another variation ofthe local independence assump- tion, which for a reflective measurement model is the ‘presumption that effect indicators are independent, ziven the (correctly specified) latent variable model. ‘The specification of correlated measurement errors relaxes this assumption and, specifically, allows for ‘omitted common causes for the pair of indicators that share an error covariance (eg, E, t_} E; for indicators X, and X,), An example of a justification for specify ing correlated errors is when two indicators share item content or a method of measurement that is unique 10 those indicators. Another example is for indicators that fare repeated-measurement variables, where correlated = 19 errors represent presumed autocorrelation (Cole et a. 2007). ‘The theoretical model implied by the specifications just described isthe domain sampling model and it as- sumes that a set of effect indicators, such a5 X,-X, in Figure 7.2(a, are internally consistent. This means that their inercorrelatons should be positive and at leat ‘moderately high in magnitude (e.. > $0). Is also assumed that that equally reliable eect indicators of the same construct ate interchangeable. That i, effet indicators can be substituted for one another without sppreciably affecting construct measurement. This a5- ‘sumption explains why it makes no sense to specify & factor with effect indicators that do not measure some- thing in common, For example, suppose that the vat- ables gender, eticity, and education are specified as effect indicators of a factor named “background” oF Some similar term, There are two problems here: Fist, gender and ethnicity are unrelated in representative samples, so these variables are not internally conss- tent: second, none ofthese indicators, such as a person's gender, isin any way “eaused” by the some underiying background” factor Formative Models ‘The assumptions of reflective measurement are not ap- propriate for some research problems, especially one ‘where composites, or “index variables" are analyzed. A “composite” is the total Score across a set of vari- ables. In this way, a composite is “caused” by its con- stituent parts, which are referred toas cause indicators ‘An example of a composite is an estimated monthly rate of inflation that takes account of price changes across different areas, such as energy, housing, food, durable goods, and so on. The meaning of a composite depends on its particular set of cause indictos that con- tribute to the total score; thus, cause indicators are not generally interchangeable, Also, cause indicators may hhave any pattern of intercorrelations: positive, negative, ‘or even clase to zero, This is because composites often reflect the contribution of multiple dimensions, albeit with a single score for each ease (e., composites are ‘ot unidimensional). There are many examples of the analysis of composites in economics and business re- search (Diamantopoulos, Riefler, & Roth, 2005). Grace (2006) describes the analysis of composites in the envi- ronmental sciences, If one believes that a set of cause indicators does nat represent all facets of the phenomenon under study,120 = then itis possible o specify a formative measurement ‘model where a latent composite has 2 disturbance, Which represents unexplained (unmeasured) variance. ‘Anexample of a formative measurement model for als tent composite with three cause indicators is presented in Figure 7.2(b. In this model, the direct effects point from the indicators, which are specified as exogenous, to the latent composite, which is endogenous and has & disturbance. The constants in the figure (1's) scale the latent composite and its disturbance. The cause indica- tors in Figure 7.2(b) are allowed to covary in any pat- tern (i, internal consistency is not required), Because cause indicators are usually specified as exogenous, their variances and covariances are not explained by 1 formative measurement model. This characteristic of formative measurement models makes it more dif- ficult to assess the validity of a set of cause indicators compared witha set of effect indicators (Bollen, 1989). In contrast, effect indicators in reflective measurement models are always endogenous, s0 the observed vari- ances and covariances of effect indicators can be com- pared against the values predicted by the model. Also, there is no error term for a cause indicator in a standard, specification of a formative model. Thus, the effect of measucement error shows up atthe construct level in- stead of being accounted for at the indicator level, as in a reflective measurement model: compare Figures 7.2(a) and 72(b). Note that the model in Figure 7.2(b) isnot identified. In order to estimate its parameter, it ‘would be necessary to embed it in a larger model. In ‘contrast, the reflective measurement model in Figure 71.2(a) is idomtificd; speciicaly, this model has no de- agrees of freedom (af = 0), but if could theoretically be analyzed as a single-factor CFA model. Formative measurement is also represented in Fig- ‘ure 7.(¢), but the composite in this model has no dis turbance. Consequently, this composite is not latent: it is instead just a total score. Grace and Bollen (2008) represent composites in model diagrams with hexa- ‘gons, which is also used in Figure 7.2(c). This is not a Standard symbol, but it does convey the fact that a com- posite with no disturbance is not latent. These same au- thors distinguish between a “fixed weights composite,” ‘here loadings (weights) are specified a priori (eg unit weighting), and an “unknown weights composite ‘where the weights are estimated with sample data. The ‘model in Figure 7.2(e) assumes an unknown weights ‘composite with a required scaling constant (1) that ap- pears along one of the paths that point from an indica- {orto the composite. Il, FUNDAMENTALS ‘The main stumbling block to analyzing measure- ‘ment models where some factors have cause indicators ‘only and the composite is latent (e.g, Figure 7.2(b) but ‘embedded in a larger model) is identification. This is because it can be difficult to specify such a model that both reflects the researcher's hypotheses and is identi- fied. The need to scale to a latent composite was men- tioned, but meeting this requirement is not dificult. ‘MacCallum and Browne (1993) noted that in order for the disturbance variance of a latent composite to be identified, the latent composite must have direct effects ‘on at least two other endogenous variables, such as en- dogenous factors with effect indicators. This require- ment is known as the “2+ emitted paths rule.” Ifa fac~ tor measured with cause indicators only emits a single pth, its disturbance variance will be not be identified, nd the analysis ofthe whole model will probably fal ‘Another requirement for models with two or more la- tent composites is that if factors measured with effect indicators only have indirect effeets on other factors that are mediated by different combinations of latent ‘composites, then some of the constituent direct effects ‘may be not be identified. Jarvis, MacKenzie, and Podsakoff (2003) advise re searchers in the consumer research area—and the rest ‘of us, too—not to automatically specify factors with effect indicators only because doing so may result in specification error, perhaps due to lack of familiar ity with formative measurement models. On the other hand, the specification of formative measurement is not 1 panacea for the reasons mentioned eavlier(identifica- ton, validity testing is more difficult), Howell, Breivik, and Wilcox (2007) were skeptical that these challenges can always be overcome and concluded that (1) forma tive measurement is not an equally attractive alterna- tive to reflective measurement and (2) researchers should try to include effect indicators whenever other indicators are specified as cause indicators ofthe same construct, but see Bagozzi (2007) and Bollen (2007) for other views. There is also a special issue in the Journal of Business Research about formative. measurement (Diamantopoutos, 2008). MIMIC Factors ‘There is actually a "compromise" between specifying thatthe indicators ofa factor are either all effect or all causal tis achieved by specifying 1 MIMIC (multiple indicators and multiple causes) factor with both effect ‘and cause indicators. There are many examples in the7. Assumptions in SEM literature ofthe analysis of structural regression models ‘with MIMIC factors. For example, Hershberger (1994) described a MIMIC depression factor with indicators ‘that represented various behaviors. Some ofthese ind ccators, such as “crying” and “feeling depressed.” were specified as effect indicators because they are symp- toms of depression. However, another indicator, “fee! ing lonely," was specified as a cause indicator. This is ‘because “feeling lonely" may be a cause of depression rather than vice versa. Bruhn, Georgi, and Hadwich (2008) describe the analysis of a MIMIC factor of cus- ‘omer equity management with latent cause indicators and manifest effect indicators. ‘A way to remedy identification problems of forma- tive measurement models isto add effect indicators for latent composites represented in the original model as ‘measured with cause indicators only; that is, specify @ MIMIC factor, For example, adding two effect indica- tors means that the formerly latent composite will emit at least two direc effects, which satisfies the 2+ emitted paths rule for the identification of the disturbances of Jatent composites—see Diamantopoulos, Riefler, and Roth (2008) for examples. However, all such respec cations require a theoretical rationale. This means that the specification of measurement as reflective, forma- tive, or MIMIC-type should reflect substantive theory for & particular research problem about the directional- ity of effects between indicators and latent variables. DATA-RELATED ASSUMPTIONS: ‘The cardinal assumption of any estimation method in ‘SEM isthatthe models correctly specified.’ Otherwise, ‘model parameters may be estimated with bias. This as- sumption is especially critical for a full-information ‘method, such as the default ML estimation, that simul- ‘taneously estimates all model parameters. The reason is the phenomenon of error propagation, which can happen in full-information methods when specifica- tion error in one part of the model affects estimates for other parameters elsewhere in the model. Suppose that ‘an error covariance for two indicators ofthe same fac- tor is incorrectly omitted. This specification error may propagate to estimation of the factor loadings for this pair of indicators. Its difficult to predict the direction ‘or magnitude of ertor propagation, but the more seri ‘ous the specification error, the more serious may be the resulting bias in other parts of the model. In computer simulation studies with incorrectly specified CFA mea- mit surement model, Bollen, Kirby, Curran, Paxton, and Chen (2007) found greater bias when using ML esti- ‘mation compared with using two-stage least squares estimation. The latter method is a partia-information ‘method that analyzes the equation of a single endoge- nous variable ata time, which may better isolate the ef- fects of error to misspecified parts of the model instead ‘of allowing them to spread to other parts. A drawback ‘of parti information methods is that there is no sta- tistical test of overall model fit (ie. there is no model chi-square); see Bollen and colleagues (2007) for more information, ‘The statistical assumptions of ML. estimation con- ‘cerning the observed variables are listed next and then discussed afterward. See also Chapter 16, this volume, by Malone and Lubansky for discussion of how to pr pare the data for analysis in SEM: 1 The observations (scores) are independent and the variables are unstandardized. 2. There are no missing values when a raw data fle is analyzed. 3. The joint distribution of the endogenous vari- ables is multivariate normal, which also implies thatthe endogenous variables are continuous 4, The exogenous variables are measured without error (Le, their score reliabilities all equal 1.00). ‘Nonindependence among scores of repested- ‘measures variables or among those from variables that share common methods of measurement can be ad: dressed through the specification of corelated error terms. In complex samples (hierarchical data sets) ‘where scores are grouped into higher-order units (eg, siblings with families) nonindependence among scores at lover levels (eg, siblings affected by family char- acteristics such as income) can be addressed through specification of a multilevel structural equation model (Kline, 2010, Chapter 12). Ifa correlation matrix is analyzed with ML estimation instead of a covariance ratrx, then values of model test statistics ad standard errors may not be correct. There are special methods for fting models to cortelaton matrices (Kline, 2010, p. 175), but default ML estimation requires unstandard- Jned variables. There is a special form of ML estima- tion for incomplete raw data iles, bu iis not available in all SEM computer tools. Otherwise, a complete raw datafile, or one with no missing values, is required1228 Distributional Assumptions Standard ML estimation makes specific distributional assumptions ofthe endogenous variables, but not of the exogenous variables, The latter can be either continu- ‘ous or categorical variables where cases are partitioned into two or more nonoverlapping groups. Group mem- bership is represented in structural models with codes, such as dummy codes or effect codes, that each rep- resent a different contrast (df= 1) between two of the ‘groups. The joint distribution of the endogenous vari- ables should be multivariate normal, which implies that (1) all univariate distributions should be normal, (2) all bivariate scatterplots are Linear, and (3) the distribution of residuals is homoscedastic. Malone and Lubansky (Chapter 16, this volume) describe how to screen raw data for multivariate normality. The normality assump- tion in ML estimation is critical Specifically, ifendog- {enous variables have severely non-normal distributions, then (1) standard errors for parameter estimates tend to bbc low, which results in rejection ofthe null hypothesis that the corresponding population parameter is zero ‘more often than is correct (Type I error rate is inflated) Also, (2) the value of the model chi-square tends 10 be too high, which results in rejection of the null hy- ppothesis that the model has perfect fit in the populs- tion more often than is correct (true models tend to be rejected to0 often). Depending on the particular pattern and severity of non-normality, however, it can also hap~ ‘pen that the value of the model chi-square can be too Jow, which means that model fit is actually worse than it appears. Because itis usually impossible to know the rection ofthe bias inthe mode! chi-square due to non- normality cis bette to thoroughly sereen the data be- forehand in order to verify distributional assumptions. ‘Transformations may be required for continuous endogenous variables with severely non-normal dis- tributions. Another option is to use robust ML estima- tion, where estimates of standard errors and model test statistics are corrected forthe degree of non-normality in the data, The corrected model chi-square generated bby most robust methods is the Satorra-Bentler statis- , Which adjusts downward the value of the standard ‘model chi-square from standard ML estimation by an amount that reflects the degree of kurtosis. Other est- ‘mation methods (ie, not ML) should be selected when some of the endogenous variables are ordinal, such as items with Likerttype scales (eg. 0 = disagree, 1 = neutral, 2 = agree). When ML estimation is used to analyze endogenous variables that are not continuous, II, FUNDAMENTALS. values of both parameter estimates and their standard errors tend to be too low (Distefano, 2002). Alter tive methods, such as robust forms of weighted least squares (WLS) estimation, are generally better choices than ML estimation for noncontinuous endogenous ‘variables. The issues just discussed imply a general sumption in SEM: The distributional characteristics of the data must match those assumed by the particular estimation method selected. Reliability Assumptions Basically any estimation method in SEM assumes that observed exogenous variables are measured without error. This includes exogenous variables in path mod- els (eg, X; and X, in Figure 71), and cause indicators in formative measurement models, for example, X,-X, in Figure 7.2(b) and (@), when such variables are not explicitly specified as indicators of underlying factors. ‘The assumption of no measurement error in exogenous variables is generally unrealistic, especially for mea sures of psychological traits, such as level of anxiety, nstead of for simple demographic variables, such as age. However, itis required forthe types of models just described because “Stand-alone” exogenous variables hhave no error terms, In contrast, endogenous variables specified as indicators of latent variables in reflective ‘measurement models have error terms, for example, E; for X, in Figure 7.2(e), and these error terms represent in part measurement error. This specification permits the estimation of parameters for latent variables, such as their variances and covariances, controlling for mea- surement error in thei indicators. ‘There is no requirement that endogenous variables in path models are measured without error, but mea- surement error in endogenous variables is manifested in their disturbances (e.g, D, for ¥; in Figure 7.1). If scores on an endogenous variable are unreliable, then its disturbance variance will be relatively large, which ‘could be confounded with omitted causes (ie, R?is re= duced). However, the potential consequences of mea- surement error in exogenous variables of path models are generally more serious than for endogenous vati- ables. Briefly, bias due to measurement error ean affect not only the path coefficient fora particular exogenous variable but also those of other exogenous variables. However, itis difficult to predict the direction and mag nitude of this error propagation. Depending on sample intercorrelations, some path coefficients for exogenous ‘variables may be biased upward (too large), but others7. Assumptions in SEM may be biased in the other direction. If the score reli abilities of exogenous variables are excellent, such as ‘rqx> 90, then the magnitude of bias may be slight, but the amount of bias increases as there is more and more ‘measurement error in exogenous variables. There is a way to respecify models, such as path models, with “stand-alone” exogenous or endogenous variables in a manner that allows direct control for measurement error. An example is presented in Figure 73. The path model in Figure 7.3(a) represents the hy- pothesis of “pure” mediation concerning the effect of ‘exogenous variable X on endogenous variable Yy; that is, ¥ is presumed to mediate all ofthe impact of X on Y, Measurement error in X will truncate the estimate ‘of the coefficient for the path X — Y, and thus increase the amount of unexplained variance (that of D,). Un- reliability in the scores of ¥, is also manifested in its (@) Path model weh pe measurement eros fj} (6) Roepeced model wih moasuremont orore (nts =n, (nts FIGURE 7.3. A path model with no measurement error for any observed variable (a) and a structural regression ‘model with a measurement error foreach observed vari- able (b). Not. ray Fys a0 Pay designate score reliability coefficients fr, respectively, variables X, ¥, and Y. = 123 disturbance D,. Because ¥, is specified asa direct cause of Y, in Figure 7:3(a), measurement error in ¥, will at tenuate the estimates of the coefficient for the path Y, ¥, and will also increase the variance of D,, ‘The model in Figure 7.3(b) isa structural regression, model, and it features: 1. The specification of X, ¥,, and Y, each as the single indicator ofa latent variable. 2, The pattern of direct effects in the structural por- tion ofthe mode! in Figure 7.3(b), or ASBOC also represents the hypothesis of “pure” media tion but now among latent variables instead of “observed variables; compare Figure 73a) and . 3. Bach observed variable in Figure 73(b) has a ‘measurement error tem. 4. The variance of each measurement error in Fi ure 7.3(b is specified to equal the product ofthe observed variance ofthe corresponding indicator and the quantity one minus the estimated score reliability for that indicator (.e. this parameter is fixed, which is required for identification). Suppose that ry = 80 for variable X in Figure 7.3(b). ‘This result implies that the proportion of variance due to random measurement error in variable X equals (1 = 80), . 20. Ifthe observed variance of X is 15.00, then the product 20 (15.00), or 3.00, estimates the error ‘variance in an unstandardized metric. The error var ‘ances for ¥, and ¥, in Figure 7.3(b) are interpreted in similar way. In the structural model of Figure 7.3(b), all path coefficients (A —> B, B > C) and disturbances, Dp, De) ace estimated explicitly controlling for mea- surement error in the indicators (Ey, Ey, and E,). See Kline (2010, pp. 276-280) for more information about the specification of single indicators with measurement crtors in structural models Exogeneity Kaplan (2004) described a way to evaluate the assump- jon of a weak form of exogeneity that is suitable for studies where all variables are measured at once (there {is no temporal precedence). Speciicaly, this assump- tion is violated if any of the following conditions does not hold: (1) the joint distribution of the endogenous124 ‘variables and the exogenous variables that are contina- ‘ous is multivariate normal; (2) all bivariate relations are linear, and (3) the distributions of regression resid- uals are all homoscedastic. Each condition just stated ccan be evalusted in data sereening. Also, each condi- tion for weak exogeneity is just an extension of a cor responding assumption of ML estimation that applies to the endogenous variables only. That is, conditions for weak exogeneity concern all the continuous vari- ables in the sample, both exogenous and endogenous, ‘but assumptions for ML estimation concern just the en- dogenous variables. Kaplan described additional tests {for stronger forms of exogeneity in panel designs in hich presumed exogenous and endogenous variables are each measured at different points in time; that is, the structural model has @ temporal structure, These tests involve checking for possible feedback effects that indicate a presumed endogenovs variable affects a pre- sumed exogenous variable (i... exogeneity is not sup. ported); see Kaplan (2004) for more information. SUMMARY Using an SEM computer tool—especislly one with 1 graphical editor that allows the user to specify the ‘model by drawing it on the sereen—itis quite easy to ‘add paths (parameters) to a structural equation model. However, one should never forget that each specitics- tion about directionality (wether structural or mes- surement) or about error covariances brings with it an assumption that requires a sound rationale. Without such a rationale, interpretation of estimates for that ps- rameter may be meaningless, Once the whole model is specified, the overarching assumption in the analysis is that (1) the model is reasonably correct and (2) the distributional characteristics of the data respect those assumed by the estimation method. Likewise, any re- specification ofthe original model inthe analysis phase requires a rationale and also implies the fundamental assumption about model correctness. Even ifthe model is ultimately retained because it is consistent with the data, never Forget thatthe model is probably one of per- haps many equivalent versions, each of which would fi the same data equally well. Also, do not falsely assume that closer to fit means closer to truth in SEM. That is, ‘models with poor fit to the data are typically respect fied by adding parameters to the model. This makes the ‘model more complex, and more complex models tend. to fit the same data better than simpler models, and Il, FUNDAMENTALS. this is true even if the more complex model is grossly misspecified. That is, model fit could improve when paths are added not because those respecifications are Correct, but simply because mew paths capitalize on ‘sample’ specific variation, Given all the assumptions of the typical SEM analy sis, the proper view that researchers should take oftheir retained model is one of skepticism. The latter should also include appreciation of the need to replicate the ‘model across independent samples and also to field test causal assumptions implied by the model that are ame- nable to the use of quasi-experimental designs among other possibilities discussed by Antonakis and col- Teagues (2010). This atitude described by the Russian novelist. Vladimir Nabokov (1899-1977) is apropos concerning the need for awareness of all assumptions ‘made in the typical application of SEM: “I am suffi- ciently proud of my knowing something to be modest about my not knowing all” NOTES 1. Boh models a Figure 71 assume tha all fects at ina, ‘nd that there are no interactive effets, ts posible in SEM, just 6 in MR, to include power or product terms in stu {oral models in onder toeatinste curvilinear orintrection et fects; ee Kline (2010, Chapter 12) and Edwards and Lambert (2007) for more information. 2, This assumption i analogous othe one for structural models that exogenous variables and omitted causes of endogenous variables are uncorelaed 3. This includes the assumption thatthe exogenous variables sand error terms are unelated. REFERENCES Antonakis, J, Bendahan,S, Jacquart,P,&Lalive,R. (2010) ‘On making causal claims: A review and recommenda. tions, Leadership Quarterly, 21(6), 1086-1120, Bagozzi, R. P. 2007). On the meaning of formative mea surement and how it differs from reflective measurement: Comment on Howell, Brevik, and Wileox (2007). Psy chological Methods, 12, 229-237 Bollen, K. A (1989), Sirutural equations with latent var ables, New York: Wiley. Bollen, K. A. (2007). Interpretational confounding is due to ‘misspecification, not t0 type af indicator: Comment on Howell, Brevik, and Wileox (2007), Peychological Meth- ‘ods, 12, 219-228. Bollea, K_A., Kinby, J.B. Curran, PJ, Paxton, P.M, & ‘Chen, F (2007). Latent variable models under misspeci-7 Assumptions in SEM fication: Two-stage least squares (TSLS) and maximum likelihood (ML) estimators. Sociological Methods and Research, 36, 48-86 Bollen, K.A. & Lennox, R. (1991). Conventional wisdom on measurement A structural equation perspective. Psycho- logical Bullein, 110, 305-314 Bruhn, M. Georgi, D, & Hadwich, K. (2008). Customer ‘equity management as formative second-order construct. ournal of Business Research, 61, 1292-1301. Cohen, J, Cohen, B, West, §.G., & Aiken, LS. (2003). Ap- plied multiple regresion/corvelation analysis for the be dhaviorl sciences Gr ed.). Mahwah, NF: EsTbaum. Cole, D. A. Ciesla, J. A. & Steiger, JH. (2007). The insii- ‘ous effects of falling to Include design-driven correlated residuals in latent-varable covariance structure analysis. Psychological Methods, 12, 381-358. Diamsantopoulos, A. (Ed). (2008). Formative indicators [Spocil issue). Journal of Business Research, 61(12). Diamantopoulos, A, Refer, P, & Roth, K. P. (2008). The problem of measurement model misspecification in behay- ‘oral and organizational esearch and some recommended Solutions. Journal of Applied Psychology, 90, 10-730. DDiamantopoulos, A Rieter, P, & Roth, K. P2008). Ad- ‘vancing formative measurement models, Journal of Busi ness Research, 61, 1203-1218. DiStefano,C. (2002). The impact of eategorization wit con firmatory factor analysis. Structural Equation Modeling 9, 327-346, Edwards, J. R, & Lambert, L, 8. (2007). Methods for inte ‘rating moderation and mediation: A general analytical framework using moderated path analysis. Psychological Methods, 12, \-22 Freedman, D. A. (1991), Statistical models and shoe lathe. ‘Socilogical Methodology, 21, 292-31. Grace, JB. (2006), Siructural equation modeling and natu ral systems. New York: Cambridge University Press Grace, J. B., & Bollen, K. A. (2008). Representing general, theoretical concepts in structural equation models: The role of composite varishles. Enviconmental and Ecologi- cal Statistics, 15, 191-213, Hershbecger,S. L. (1994). The spocitication of equivalent ‘models before the collection of data In A. von Eye & C CC. Clogg (Eis), Latent variables analyse (pp. 68-105), ‘Thousand Oaks, CA: Sage. Houghton, J. D, & Jinkerson, D. L. (2007). Constructive ‘thought strategies and job stistaction: A preliminary ex. amination. Journal of Business Paychology, 22, 48-83 = 125 Howell, R.D., Breivik, E, & Wileox J.B. 2007). Revon- ‘sidering formative measurement, Psycholosial Methods, 12, 205-218, James, L. R., Mulaik, 8. A, & Brett, J. M, (1982). Causal ‘analysts: Models, assumptions and dara. Beverly Hills, CA: Sage. Jarvis, CB, MacKenzie, S.B., & Podsakoff, P.M. (2003) ‘Arita review of construct indicators and measurement ‘model misspecification in marketing and consumer re- search, Journal of Consumer Research, 30, 199-218. ‘Kaplan, D. (2004), On exogeneity. In D. W. Kaplan (Ed), The ‘Sage handbook of quantitative methodology inthe social sciences (pp. 407-421), Newbury Park, CA: Sage ‘Kaplan, D.,Harik,P, & Hotchkiss, L. 2061) -Cross-setional ‘estimation of dynamic structural equation models in lisequilibrium. In R. Cudeck, S. Du Tot, & D. Sérbom (Eds), Structural equation modeling: Present and future: A Festschrift in honor of Kari Jéreskog (pp. 315-239), Lincolnwood, TL: Seiemile Sofware International Kenny, D. A. (2008). Mediation. Retrieved April 20, 2008, from davidakenny.net/em/mediat.hum. Kesolman, H. 1, Hubert, C.J, Lix, L. M, Olsjik, S,Crib- bie, R.A, Donahue, B, e al. (1998). Statistical practices ‘of education researchers: An analysis of the ANOVA, MANOVA. and ANCOVA analyses. Review of Educa onal Research, 68, 380-368. Kline, R. B. (2010). Principles and practice of structural ‘equation modeling (3d ed). New York: Guilford Press. MacCallum, R-C., & Browne, M. W. (1993). The use of cas indicators in covariance structure models: Some prac cal issues. Psychological Bulletin, 114, $33-S4l. Pearl, J. 2000). Causality: Models, reasoning, and iner- ence. New Yark: Cambridge University Pres. Peat, J (2009). Causality: Models, reasoning. and inference (Qnd ed). New York: Cambridge Univesity Press. Sava, F. A. 2002), Causes and effects of teacher conflict- inducing attitudes towards pupils: A path analysis model Teaching and Teacher Education, 18, 1007-1021 Shah, R. & Goldstein, S.M. (2006). Use of structural equa- ‘tion modalng in operations management research: Look: ing back and forward. Journal of Operations Manage- ‘ment, 24, 148-168 ‘Wilcox, R. R. (1998). How many discoveries have been lost by ignoring modern statistical methods? American Psy- chologist, 53, 300-314. Wirght . (1923), The theory of path coefficients: A reply to ‘Niles’ erin, Gneties, 20, 239-255