G Lect 31 G Lecture 3 SEM Model notation Review of mediation Estimating SEM models Moderation
G Lect 32 SEM Notation for LISREL (Joreskog) Lisrel's notation is used by authors such as Bollen Y2Y2 Y1Y1 X1X1
G Lect 33 SEM Notation for EQS (Bentler) EQS does not name coefficients. It also does not distinguish between exogenous and endogenous variables. EE V3V3 V2V2 V1V1 EE
G Lect 34 SEM Notation for AMOS (Arbuckle) AMOS does not use syntax, and it has no formal equations. It is graphically based, with user-designed variables. EE Fa Fiz Foo EE
G Lect 35 Revisiting Mediation We wish to “explain” modeled path c: with: Total mediation models identify instruments for sophisticated structural equation models. Y eyey X M c' b eMeM a X c e Y
G Lect 36 Mediation: A theory approach Sometimes it is possible to argue on theoretical grounds that Z is prior to X and Y X is prior to Y The effect of Z on Y is completely accounted for by the indirect path through X. This is an example of total mediation If is fixed to zero, then Model 3 is no longer saturated. Question of fit becomes informative Total mediation requires strong theory
G Lect 37 Nonrecursive models All the models we have considered are recursive. The causal effects move from one side of the diagram to the other. A Nonrecursive model has loops or feedback. X1X1 Y1Y1 X2X2 Y2Y2 e1e1 e2e2
G Lect 38 Equations Implied Assume all variables centered Predicting Y 1 : Y 1 = b 1 Y 2 + g 1 X 1 + e 1 Predicting Y 2 : Y 2 = b 2 Y 1 + g 2 X 2 + e 2 Note that the error terms cannot be assumed to be independent. There are no direct paths between X 1 and Y 2, or between X 2 and Y 1.
G Lect 39 Example Let X 1 be college aspirations of parents of adolescent 1 and X 2 be aspiration of parents of adolescent 2. Youth 1 aspirations (Y 1 ) are affected by her parents and her best friend (Y 2 ), and Youth 2 has the reciprocal pattern. This model is identified because it assumes that the effect of X 1 on Y 2 is completely mediated by Y 1. Special estimation methods are needed. OLS gives biased results.
G Lect 310 SEM and OLS Regression SEM models and multiple regression often lead to the same results When variables are all manifest When models are recursive The challenges of interpreting direct and indirect paths are the same in SEM and OLS multiple regression SEM estimates parameters by fitting the covariance matrix of both IVs and DVs
G Lect 311 Matrix Notation for SEM Consider LISREL notation for this model: Y2Y2 Y1Y1 X2X2 X1X1
G Lect 312 More Matrix Notation The matrix formulation also requires that the variance/covariance of X be specified Sometimes is used, sometimes XX. The variance/covariance of is also specified Conventionally this is called . When designing structural models, the elements of and can either be estimated or fixed to some (assumed) constants.
G Lect 313 Basic estimation strategy Compute sample variance covariance matrix of all variables in model Call this S Determine which elements of model are fixed and which are to be estimated. Arrange the parameters to be estimated in vector . Depending on which values of are assumed, the fitted covariance matrix ( ) has different values Choose values of that make the S and as close as possible according using a fitting rule
G Lect 314 Estimates Require Identified Model An underidentified model is one that has more parameters than pieces of relevant information in S. The model should always have where t is the number of parameters, p is the number of Y variables and q is the number of X variables Necessary but not sufficient condition
G Lect 315 Other identification rules Recursive models will be identified Bollen and others describe formal identification rules for nonrecursive models Rules involve expressing parameters as a function of elements of S. Informal evidence can be obtained from checking if estimation routine converges However, a model may not converge because of empirical problems, or poor start values
G Lect 316 Review of Expectations The multivariate expectations Var( X + k* 1 ) = Var(k* X ) = k 2 Let C T be a matrix of constants. C T X=W are linear combinations of the X's. Var(W) = C T Var(X) C = C T C This is a matrix
G Lect 317 Multivariate Expectations In the multivariate case Var(X) is a matrix V(X)=E[(X- ) (X- ) T ]
G Lect 318 Expressing If Then We get by specifying the model details. We also consider Cov(XY)
G Lect 319 Estimation Fitting Functions ML minimizes ULS minimizes GLS minimizes
G Lect 320 Excel Example
G Lect 321 Choosing between methods ML and GLS are scale free Results in inches can be transformed to results in feet ULS is not scale free All methods are consistent ML technically assumes multivariate normality, but it actually is related to GLS, which does not Parameter estimates are less problematic than standard error estimates
G Lect 322 Moderation Baron and Kenny (1986) make it clear that mediation is not the only way to think of causal stages A treatment Z may enable an effect of X on Y For Z=1 X has effect on Y For Z=0 X has no effect on Y When effect of X varies with level of Z we say the effect is Moderated SEM methods do not naturally incorporate moderation models
G Lect 323 Moderation, continued In multiple regression we add nonlinear (e.g. multiplicative) terms to linear model Covariance matrix is expanded Distribution of sample covariance matrix is more complex SEM ability to represent latent variables in interactions is limited Easiest case is when moderator is discrete
G Lect 324 Suppose that X is perceived efficacy of a participant and Y is a measure of influence at a later time. Suppose S is a measure of perceived status. Perceived status might moderate the effect of efficacy on influence. Two ways to show this: Equation: Y=b 0 +b 1 X+b 2 S+b 3 (X*S)+e Path Diagrams of Moderation X S Y e For S low: For S high: XY e XY e (+) +