General Structural Equations - PowerPoint PPT Presentation

About This Presentation
Title:

General Structural Equations

Description:

General Structural Equations Week 1 Class #5 Today: More on estimation More on block tests Out of range solutions: what they mean & how to deal with them Higher order ... – PowerPoint PPT presentation

Number of Views:71
Avg rating:3.0/5.0
Slides: 44
Provided by: Douglas250
Category:

less

Transcript and Presenter's Notes

Title: General Structural Equations


1
General Structural Equations
  • Week 1 Class 5

2
Today
  • More on estimation
  • More on block tests
  • Out of range solutions what they mean how to
    deal with them
  • Higher order latent variable models
  • When a latent variable has other latent variables
    as indicators
  • Using the PRELIS program to generate covariance
    matrices for LISREL
  • A quick look at SAS-CALIS
  • Item parcels (a controversy?)
  • An extended discussion of an example set
  • If time permits re-expressing latent variable
    structural equation models in matrix terms

3
First computer assignment
  • Due Tuesday
  • Short answer responses (submit some of the
    programs if AMOS, diagrams or diskette)
  • REQUIRED for credit, letter with grade
    participants
  • Choice of AMOS or SIMPLIS. If using SIMPLIS, a
    system (.dsf) file has already been created.

4
PRELIS demonstrationReview of SIMPLIS programs
(ReligSexMor problem)
5
Estimation (notes)
  • Parameter estimates are obtained through
    iterative methods
  • Start with start values
  • Could be user guesses (early versions of
    LISREL)
  • Could use some other single-step estimation
    method (eg 2SLS)
  • Use start values to calculate reproduced
    covariance matrix

6
Estimation (notes)
  • Start with start values
  • Could be user guesses (early versions of
    LISREL)
  • Could use some other single-step estimation
    method (eg 2SLS)
  • Use start values to calculate reproduced
    covariance matrix
  • Calculate first order derivatives for each free
    parameter
  • Will tell us for any given parameter whether next
    iteration value should be higher or lower e.g.,
    positive derivative means value is too high
  • Optionally, calculate second order derivatives
  • Computationally intensive (usually)
  • Trade off between extra effort at re-calculation
    (sometimes, matrix is merely updated with an
    approximation and only fully re-calculated every
    X iterations) and precision
  • Sometimes, programs unable to calculate matrix of
    2nd order derivatives with given start values and
    use Steepest Descent methods (esp. initially)

7
Estimation (notes)
  • Convergence declared when estimates of fit
    function become sufficiently similar from one
    iteration to next and/or parameter estimates
    dont differ by more than x (convergence
    criterion)
  • Occasionally, a model will not converge
  • Check the model to make sure there are no
    implausible paths, model is identified
  • Check to make sure there arent any ve error
    variances how to deal with these will be
    discussed later
  • If otherwise OK except for non-convergence, ask
    for more iterations (most software will allow
    this)

8
Problems
  • The negative error variance.
  • In theory, it is impossible for a variance to be
    negative
  • But, SEM models can be estimated where the fit
    function minimum occurs when one of the
    parameters is improperly negative.

9
Problems
  • The negative error variance.
  • Main reasons
  • Improperly specified model (e.g., missing a
    parameter that would improve the fit
    considerably)
  • Frequently occurs in models with LVs with 2
    indicators
  • Sampling distribution (the real parameter is
    positive in the population, but in our sample, it
    ends up being ve)
  • More likely to happen in smaller samples

10
Problems
  • The negative error variance.
  • How the software responds to it
  • Allow the parameter to go negative, perhaps
    providing a warning (matrix not
    positive-definite)
  • Allow a finite number of further iterations, and
    if the parameter doesnt become positive, stop
    the iteration process and generate a warning
    message (LISREL default)
  • Impose an inequality constraint

11
Problems
  • The negative error variance.
  • How the investigator should respond to it
  • Check the model carefully. Are there any
    inappropriate paths? Major (likely) paths that
    are missing? In 2-indicator LV models, can a 3rd
    indicator be found?
  • Check the significance of the ve parameter (if
    necessary, re-run the model over-riding the
    defaults so that the model is allowed to reach
    convergence)
  • Constrain the variance parameter to zero or some
    small value.

12
Higher order models
  • 2nd order models where the indicators for a
    latent variable are themselves latent variables
  • many of the same principles apply treat 1st
    level latent variables as indicators of 2nd level
    latent variables
  • will need reference indicator, for example
  • first-level latent variables must be
    sufficiently correlated (or model will not work)

13
(No Transcript)
14
Higher order models start with more modest
ambitions and test a correlated LV
If correlations among LVs are low, it may not be
reasonable or even possible to estimate a
higher-order model
Only two indicators? Same issues as with
lower-level LVs.
15
Is this model adequate?
16
Or is something like this required?
17
Block tests
SAME ISSUES WHETHER L1-L3 are LVs, single
indicator exogenous vars or dummy variables
18
Block Tests
  • Matrix
  • L1 L2 L3 exogenous
  • L4 b1 b2 b3 (Model 1)
  • L5 b4 b5 b6
  • Test of whether L1 has effect on endog.
    variables
  • Model 2, as above but b10 b40
  • Model 3, b1 through b6 0
  • Model 4, b1?0, b4 ?0, b2 b3 b5 b6 all 0t
  • Test of equation with L4 dependent
  • Model 5, as above but b10, b20, b30
  • Compare models using chi-square difference (df
    difference in of degrees of freedom between
    models)

19
Tests involving dummy (exogenous) variable
contrasts
  • Example
  • Variable religion
  • Categories Protestant D11
  • Fundamentalist Prot D21
  • Catholic D31
  • Muslim D41
  • Atheist/agnostic reference
  • (D1D2D3D40)

20
Tests involving dummy (exogenous) variable
contrasts
  • Example
  • Variable religion
  • Categories Protestant D11
  • Fundamentalist Prot D21
  • Catholic D31
  • Muslim D41
  • Atheist/agnostic reference
  • (D1D2D3D40)
  • Important dummy variables are not (usually)
    orthogonal. Make sure to allow for covariances
    among them exception orthogonally coded or
    effects coded in balanced designs
  • Coefficients
  • Protestant b1 Tests Protestant vs. Atheist
  • Fund. Prot b2 Tests Fund. Prot. Vs. Atheist
  • Catholic b3 Tests Cath. Vs. atheist
  • Muslim b4 Tests Muslim vs. atheist
  • Other tests
  • Protestant vs. Catholic? Run a new model with
    b1b3, compare chi-sq.
  • Protestants AND fund. Prot. TOGETHER vs.
    Catholic?
  • Model 1 b1b2

21
Item parcels
Versus
Assume 0 error variance or estimate from
reliability coeff.
X9

Add scores of X1X2X3X4X5X6X7X8 to get an
item parcel
22
Item parcels
  • Less extreme

23
Reasons for using item parcels
  • With single indicator models, will get fits that
    are very close to perfect (bad reason!)
  • Individual indicators may be non-normally
    distributed (e.g., 4-category, 5-category
    attitude scale items tend to be kurtotic)
    summing indicators will often help
  • Individual indicators may be extreme (e.g,
    dichotomies, tricotomies)
  • The model may be monstrous (dozens of
    indicators per construct hundreds of variables
    in model) with a lot of somewhat redundant
    information (alternative would be to randomly
    select indicators, but why throw away data?)

24
Reasons for NOT using parcels
  • One of the big advantages of LV SEM models is
    discarded (at least in extreme cases where items
    reduced to a single item)
  • Hypothesized pattern of inter-relationships among
    indicators may be incorrect (suggestion test by
    running model on items to be parceled, if
    possible)
  • Even if items internally consistent, assumes
    internal consistency will imply consistency with
    respect to other variables in the model

25
LV Structural Equation Models in Matrix terms
  • Thus far, our work has involved scalar
    equations.
  • one equation at a time
  • Specify a model (e.g, with software) by writing
    these equations out, one line per equation

26
Matrix form
  • We can represent the previous 2 equations in
    matrix form

Matrix Form (single, double subscript)
27
There are other matrices in this model
Variance-covariance matrix of error terms (es)
28
(other matrices, continued)
Variance covariance matrix of exogenous
(manifest) variables
29
Two scalar equations re-written
scalar
Matrix
Contents of matrices
30
More generic form (combines all exogenous
variables into single matrix)
More generic
Where E1 ? X1, E2 ? X2 and E3 ? X3
31
More generic form
All exogenous variables part of a single
variance-covariance matrix
32
Reproduced covariances (the formula in matrix
terms)
T above elements of which are called ?theta
is not the same as ? in S(?). Latter refers to
all parameters in a model. Theta above refers to
elements in the variance-covariance matrix of
errors/exogenous variables.
33
A simple model
B
Continued..
34
Reproduced covariances (observed variable model
without latent variables)
35
(No Transcript)
36
(proof of inverse quick aside)
37
(No Transcript)
38
Measurement (factor) model
39
(No Transcript)
40
Alternative notation systems for coefficients
41
(No Transcript)
42
(No Transcript)
43
(No Transcript)
Write a Comment
User Comments (0)
About PowerShow.com