1 / 54

General Structural Equation (LISREL) Models

General Structural Equation (LISREL) Models. Week #2 Class #1. Today’s class:. Standardization A review: the “metric” of latent variables Relationship with reference indicator Variances of latent variables Interpreting coefficients Other metric/scaling issues

gporter
Télécharger la présentation

General Structural Equation (LISREL) Models

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. General Structural Equation (LISREL) Models Week #2 Class #1

  2. Today’s class: • Standardization • A review: the “metric” of latent variables • Relationship with reference indicator • Variances of latent variables • Interpreting coefficients • Other metric/scaling issues • Variables with very large or very small variances • Quick introduction to programming in SAS and EQS • Structural Equation Models in Matrix form

  3. Standardization • bstandardized bz= b * {SQRT[var(x)] /SQRT[ var(y)]} • In a measurement model, • bz = b * ( SQRT variance of latent variable ) SQRT variance of manifest variable • In a construct equation model • bz = b * (SQRT variance of exogenous LV / SQRT variance of endogenous LV)

  4. Standardization • In a measurement model, • bz = b * ( SQRT variance of latent variable ) sqrt variance of manifest variable • What LISREL calls “standardized” omits the division by the manifest variable • Look for “COMPLETELY STANDARDIZED” solution • Standardized coefficients are useful for comparing the size of effects of different variables in a single model • As with regular regression, not so useful comparing across samples

  5. Standardization • Standardized coefficients are useful for comparing the size of effects of different variables in a single model • As with regular regression, not so useful comparing across samples • Generally not appropriate for dummy variables • Especially inappropriate for dummy variables where # of categories for a categorical variable > 2

  6. Standardization • Generally not appropriate for dummy variables • Especially inappropriate for dummy variables where # of categories for a categorical variable > 2 • Can calculate R-square for the entire variable

  7. R-square for categorical variables Country with 4 categories: USA; Canada; France; Germany (Germ=ref) R2 explained for country: 1. Get R2 for model shown 2. Subtract R2 for model with B1=0, B2=0, B3=0

  8. Variances of latent variables What is the variance of LV1? X1 = 1.0 LV1 + e1 VAR(X1) = COV(LV1,LV1) + 2*COV(LV1,e1) + COV(e1,e1) = VAR(L1) + VAR(e1) VAR(L1) = VAR(X1) – VAR(e1) If VAR(e1) is small, VAR(L1) is ≈ VAR(X1) Otherwise, VAR(L1) < VAR(X1) Alternative: fix VAR(L1) = 1.0

  9. Variances of latent variables In this model, we can no longer fix VAR(LV1) = 1.0 We could fix VAR(d1) = 1.0 If so, VAR(LV1) = b12 VAR(X4) + VAR(d1) (WE could impose a complicated constraint: VAR(d1) = 1 – ( b12 * VAR(X4)) b12, VAR(X4) are model parameters Complex constraints of this sort not possible in AMOS

  10. Variances of latent variables We could fix VAR(d1) = 1.0 If so, VAR(LV1) = b12 VAR(X4) + VAR(d1) (WE could impose a complicated constraint: VAR(d1) = 1 – ( b12 * VAR(X4)) b12, VAR(X4) are model parameters Complex constraints of this sort not possible in AMOS Also: more complicated with multiple exogenous variables, even more complicated if X4-type variables are latent var’s

  11. Variances of latent variables We could fix VAR(d1) = 1.0 If so, VAR(LV1) = b12 VAR(X4) + VAR(d1) (WE could impose a complicated constraint: VAR(d1) = 1 – ( b12 * VAR(X4)) b12, VAR(X4) are model parameters Complex constraints of this sort not possible in AMOS Also: more complicated with multiple exogenous variables, even more complicated if X4-type variables are latent var’s

  12. How do we interpret b1?

  13. X1 = Avg. church attendance per month X2 = 4 point scale, 4=very religious through 1 = not at all X3 = 10 point scale, importance of religion in my life X4 = 10 point scale, 1=abortion never justified 10 = abortion always justified X5 = 4 point scale, Abortion a matter of woman’s choice: 4=agree strongly through 1 = disagree strongly X6 = 5 point scale, Abortion should never be allowed 5=agree strongly through 1 = disagree strongly (3=“neutral”)

  14. X1 = Avg. church attendance per month * X2 = 4 point scale, 4=very religious through 1 = not at all X3 = 10 point scale, importance of religion in my life X4 = 10 point scale, 1=abortion never justified 10 = abortion always justified * X5 = 4 point scale, Abortion a matter of woman’s choice: 4=agree strongly through 1 = disagree strongly X6 = 5 point scale, Abortion should never be allowed 5=agree strongly through 1 = disagree strongly (3=“neutral”)

  15. X1 = Avg. church attendance per month * X4 = 10 point scale, 1=abortion never justified 10 = abortion always justified * Interpretation: 1 unit increase in church attendance per month leads to b1 increase In abortion support along 10-point scale

  16. Interpretation: 1 unit increase in church attendance per month leads to b1 increase In abortion support along 10-point scale Because the variance of the dependent variable is likely to be (much) higher than the variance of the independent variable, we expect large values for b1 (could be >1 if effect is strong). If this were reversed, we would have small values of b1, even if there is a strong effect.

  17. If X1 is measured on a 4-point scale, Variance of AbortAtt might be quite small (e.g., .5 or .3 or less) Variance of age can be quite large (e.g., 300, 400).  this implies that b2 can be very small, even if effect is large (standardized b2 not affected) Could scale down variance of Age (in SPSS: Age = Age/10)

  18. SAS, EQS programming. • The basics: • Specify equations • Specify “variances” (list variables for which variances are model parameters) • Specify “covariances (list covariances among exogenous variables, if any)

  19. SAS, EQS programming. Sample SAS program: proc calis m=ml mod data=matrix1 cov; lineqs v275 = f1 + e1, v276 = b1 f1 + e2, v277 = b2 f1 + e3, v278 = b3 f1 + e4, v279 = b4 f1 + e5, v280 = b5 f1 + e6, v281 = b6 f1 + e7, v282 = b7 f1 + e8, v283 = b8 f1 + e9; std f1 =vf1, e1=ve1, e2=ve2, e3=ve3, e4=ve4, e5=ve5, e6=ve6, e7=ve7, e8=ve8 , e9=ve9; Run; In SAS, all parameters must be named. Can’t just say “variance of f1 is free” – must give it a parameter name

  20. EQS programming Sample program: /title any title /spec cas=342; var=8; ma=cov; an=cov; da=‘c:\data1.cov’; fo=‘(8f10.2)’; /equ v1 = *F1+ e1; v2 = 1.0 F1 + e2; v3 = *F1 + e3; V4 = *F1 + *F2 + e4; V5 = 1.0 F2 + e5; V6 = *F2 + e6; v7 = *F2 + e7; V8 = *F2 + e8; /var f1=*; f2=*; e1 to e8 = *; /cov f1,f2=* /lmtest /end

  21. LV Structural Equation Models in Matrix terms (continued) From last week

  22. Two scalar equations re-written scalar Matrix Contents of matrices

  23. Reproduced covariances (the formula in matrix terms) Θ above – elements of which are called θ[theta] is not the same as θ in Σ(θ). Latter refers to all parameters in a model. Theta above refers to elements in the variance-covariance matrix of errors/exogenous variables.

  24. A simple model: B Continued……..

  25. Reproduced covariances (observed variable model without latent variables)

  26. (proof of inverse: quick aside)

  27. Measurement (“factor”) model

  28. Alternative notation systems for coefficients:

  29. Matrix Form Latent variables only: PHI (Φ) GAMMA ( Γ) PSI (Ψ)

  30. Matrix form: LISREL matrices LISREL calls the LVs ξ (KSI) or η (ETA), depending on whether they are exogenous or endogenous

  31. Matrix form: LISREL matrices The GAMMA (Γ) matrix is for paths from KSI (ξ ) latent variables to ETA (η ) latent variables. KSI variables must be completely exogenous. In the model shown below, ETA-1 (η1 ) is exogenous with respect to ETA-2 (η2 ) but endogenous with respect to KSI-1 (ξ1 ) . • It is therefore endogenous -- an ETA and not a KSI variable.

  32. Matrix form: LISREL matrices Paths from one ETA (latent) variable to another are represented by elements in the BETA ( Β ) matrix. GAMMA 2 x 2 BETA ( 2 x 2)

  33. Matrix form: LISREL matrices BETA will be sub-diagonal if the model is recursive:

  34. Matrix form: LISREL matrices BETA will be sub-diagonal if the model is recursive: Can be re-expressed as:

  35. Matrix form: LISREL matrices BETA will be sub-diagonal if the model is recursive, but above-diagonals will be used if the model is non-recursive:

  36. Matrix form: LISREL matricesMEASUREMENT MODEL MATRICES From above: Two types of latent variables: ETA (η ) endogenous KSI (ξ ) completely exogenous MEASUREMENT MODEL FOR KSI VARIABLES: Manifest variables: X’s Measurement errors: DELTA ( δ) Coefficients in measurement equations: LAMBDA ( λ ) Sample equation: X1 = λ1ξ1+ δ1

  37. Matrix form: LISREL MEASUREMENT MODEL MATRICES Manifest variables: X’s Measurement errors: DELTA ( δ) Coefficients in measurement equations: LAMBDA ( λ ) Sample equation: X1 = λ1ξ1+ δ1 MATRICES: LAMBDA-x THETA-DELTA PHI

  38. Matrix form: LISREL MEASUREMENT MODEL MATRICES A slightly more complex example:

  39. Matrix form: LISREL MEASUREMENT MODEL MATRICES Labeling shown here applies ONLY if this matrix is specified as “diagonal” Otherwise, the elements would be: Theta-delta 1, 3, 6, 10, 15. OR, using double-subscript notation: Theta-delta 1,1 Theta-delta 2,2 Theta-delta 3,3 Etc.

  40. Matrix form: LISREL MEASUREMENT MODEL MATRICES While this numbering is common in some journal articles, the LISREL program itself does not use it. Two subscript notations possible: Single subscript Double subscript

  41. Matrix form: LISREL MEASUREMENT MODEL MATRICES Models with correlated measurement errors:

  42. Matrix form: LISREL MEASUREMENT MODEL MATRICES Measurement models for endogenous latent variables (ETA) are similar: • Manifest variables are Ys • Measurement error terms: EPSILON ( ε ) • Coefficients in measurement equations: LAMBDA (λ) • same as KSI/X side • to differentiate, will sometimes refer to LAMBDAs as Lambda-Y (vs. Lambda-X) • Equations • Y1 = λ1η1+ ε1

  43. Matrix form: LISREL MEASUREMENT MODEL MATRICES Measurement models for endogenous latent variables (ETA) are similar:

  44. LISREL MATRIX FORM An Example:

More Related