1 / 32

Structural Equation Modeling

Structural Equation Modeling. Continued: Lecture 2 Psy 524 Ainsworth. Covariance Algebra. Underlying parameters in SEM Regression Coefficients Variances and Covariances A hypothesized model is used to estimate these parameters for the population (assuming the model is true)

scott
Download Presentation

Structural Equation Modeling

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Structural Equation Modeling Continued: Lecture 2 Psy 524 Ainsworth

  2. Covariance Algebra • Underlying parameters in SEM • Regression Coefficients • Variances and Covariances • A hypothesized model is used to estimate these parameters for the population (assuming the model is true) • The parameter estimates are used to create a hypothesized variance/covariance matrix • The sample VC matrix is compared to the estimated VC

  3. Covariance Algebra • Three basic rules of covariance algebra • COV(c, X1) = 0 • COV(cX1, X2) = c * COV(X1, X2) • COV(X1 + X2, X3) = COV(X1, X3) + COV(X2, X3) • Reminder: Regular old fashioned regression • Y = bX + a

  4. Example Model

  5. Covariance Algebra • In covariance structure intercept is not used • Given the example model here are the equations • Y1=g11X1 + e1 • Here Y1 is predicted by X1 and e1 only • X1 is exogenous so g is used to indicate the weight • Y2=b21Y1 + g21X1 + e2 • Y2 is predicted by X1 and Y1 (plus error) • Y1 is endogenous so the weight is indicated by b • The two different weights help to indicate what type of relationship (DV, IV or DV, DV)

  6. Covariance Algebra Estimated Covariance given the model • Cov(X1,Y1) the simple path • Since Y1=g11X1 + e1 we can substitute • COV(X1,Y1) = Cov(X1, g11X1 + e1) • Using rule 3 we distribute: • COV(X1, g11X1 + e1) = Cov(X1, g11X1)+ (X1, e1) • By definition of regression (X1, e1) = 0 • So this simplifies to: Cov(X1, g11X1 + e1) = Cov(X1, g11X1) • Using rule 2 we can pull g11 out: • COV(X1,Y1) = g11Cov(X1, X1) • COV(X1, X1) is the variance of X1 • So, Cov(X1,Y1) = g11sx1x1

  7. Covariance Algebra Estimated Covariance given the model • COV(Y1,Y2) the complex path • Substituting the equations in for Y1 and Y2 • COV(Y1,Y2) = COV(g11X1 + e1, b21Y1 + g21X1 + e2) • Distributing all of the pieces • COV(Y1,Y2) = COV(g11X1,b21Y1) + COV(g11X1, g21X1) + COV(g11X1, e2) + COV(e1, b21Y1) + COV(e1, g21X1) + COV(e1, e2) • Nothing should correlate with e so they all drop out • COV(Y1,Y2) = COV(g11X1,b21Y1) + COV(g11X1, g21X1) • Rearranging: COV(Y1,Y2) = g11b21COV(X1,Y1) + g11g21COV(X1,X1) COV(Y1,Y2) = g11b21sx1y1 + g11g21sx1x1

  8. COV(X1,Y1) = g11s2x1 COV(Y1,Y2) = g11b21sx1y1 + g11g21s2X1 Back to the path model

  9. Example

  10. Example

  11. SEM models • All of the relationships in the model are translatable into equations • Analysis in SEM is preceded by specifying a model like in the previous slide • This model is used to create an estimated covariance matrix

  12. SEM models • The goal is to specify a model with an estimated covariance matrix that is not significantly different from the sample covariance matrix • CFA differs from EFA in that the difference can be tested using a Chi-square test • If ML methods are used in EFA a chi-square test can be estimated as well

  13. Model Specification • Bentler-Weeks model • Matrices • B – Beta matrix, matrix of regression coefficients of DVs predicting other DVs • g – Gamma matrix, matrix of regression coefficients of DVs predicted by IVs • F – phi matrix, matrix of covariances among the IVs •  - eta matrix, vector of DVs •  - xi matrix, vector of IVs • Bentler-Weeks regression model

  14. Model Specification • Bentler-Weeks model • Only independent variables have covariances (phi matrix) • This includes everything with a single headed arrow pointing away from it (e.g. E’s, F’s, V’s, D’s, etc.) • The estimated parameters are in the B, g and F matrices

  15. Model Specification • Bentler-Weeks model matrix form

  16. Diagram again

  17. Model Specification • Phi Matrix • It is here that other covariances can be specified

  18. Model Estimation • Model estimation in SEM requires start values • There are methods for generating good start values • Good means less iterations needed to estimate the model • Just rely on EQS or other programs to generate them for you (they do a pretty good job)

  19. Model Estimation • These start values are used in the first round of model estimation by inserting them into the Bentler-Weeks model matrices (indicated by a “hat”) in place of the *s • Using EQS we get start values for the example

  20. Model Estimation

  21. Model Estimation • Selection Matrices (G) • These are used to pull apart variables to use them in matrix equations to estimate the covariances • So that Y = Gy *  and Y is only measured dependent variables

  22. Model Estimation • Gx = [1 0 0 0 0 0 0] so that: X = Gx *  and X is (are) the measured variable(s) • Rewriting the matrix equation to solve for  we get: =(I-)-1 • This expresses the DVs as linear combinations of the IVs

  23. Model Estimation • The estimated population covariance matrix for the DVs is found by:

  24. Model Estimation • Estimated covariance matrix between DVs and IVs

  25. The covariance(s) between IVs is estimated by : Model Estimation

  26. Model Estimation • Programs like EQS usually estimate all parameters simultaneously giving an estimated covariance matrix sigma-hat • S is the sample covariance matrix • The residual matrix is found by subtracting sigma-hat from S • This whole process is iterative so that after the first iteration the values of the hat matrices output are input as new starting values

  27. Model Estimation • Iterations continue until the function (usually ML) is minimized, this is called convergence • After five iterations the residual matrix in the example is:

  28. Model Estimation • When the model converges you get estimates of all of the parameters and a converged residual matrix

  29. Model Estimation

  30. Model Evaluation • 2 is based in the function minimum (from ML estimation) when the model converges • The minimum value is multiplied by N – 1 • From EQS: (.08924)(99) = 8.835 • The DFs are calculated as the difference between the number of data points (p(p + 1)/2) and the number of parameters estimated • In the example 5(6)/2 = 15, and there are 11 estimates leaving 4 degrees of freedom

  31. Model Estimation • 2(4) = 8.835, p = .065 • The goal in SEM is to get a non-significant chi-square because that means there is little difference between the hypothesized and sample covariance matrices • Even with non-significant model you need to test significance of predictors • Each parameter is divided by its SE to get a Z-score which can be evaluated • SE values are best left to EQS to estimate

  32. Final example model with (unstandardized) and standardized values Final Model

More Related