Académique Documents
Professionnel Documents
Culture Documents
Estimation
In this step, start values of the free parameters are chosen in order to
generate an estimated population covariance matrix, E(u), from the model.
Start values can be chosen by the researcher from prior information, by
computer programs used to build SEMs , or from multiple regression analysis
(See Ullman 1996 and Hoyle 1995 for more start value choices and further
discussion).
The goal of estimation is to produce a E(u) that converges upon the observed
population covariance matrix, S, with the residual matrix (the difference
between E(u) and S) being minimized.
Various methods can be used to generate E(u). Choice of method is guided
by characteristics of the data including sample size and distribution. Most
processes used are iterative. The general form of the minimization function
is:
Q = (s - o(u))W(s - o(u))
where, s = vector containing the variances and covariances of the observed
variables
o(u) = vector containing corresponding variances and covariances as
predicted by the model
W = weight matrix.(Some authors refer to Q as F).
The weight matrix, W, in the function above, corresponds to the estimation
method chosen. W is chosen to minimize Q, and Q(N-1) gives the fitting
function, in most cases a X2-distibuted statistic. The performance of the X2 is
affected by sample size, error distribution, factor distribution, and the
assumption that factors and errors are independent (Ullman 1996).
Some of the most commonly used
estimation methods are:
Generalized Least Squares (GLS)
FGLS = tr[([S - E(u)]W-1)2]
tr = trace operator, takes sum of elements on main diagonal of matrix
W-1 = optimal weight matrix, must be selected by researcher (most common
choice is S-1)
Maximum Likelihood (MLE) most often used.
FML = log|E| - log|S| + tr(SE-1) - p
W = E-1 and p = number of measured variables
Asymptotically Distribution Free (ADF) Estimator
FADF = [S - o(u)]W-1[S - o(u)]
W, in this function, contains elements that take into account kurtosis.
Ullman (1996) and Hoyle (1995) discuss the
advantages and limitations of the above estimators.
ML (less problematic) and GLS are useful for normally
distributed, continuous data when factors and errors are
independent. Refer to handout
ADF is useful for analyzing categorical data, but is shown only
to work well with sample sizes above 2,500. Refer to handout
What if the data is categorical and the sample size is not so
large? Use measurement with 4 or more categories.
Ullman indicates that the best estimator for non-normally
distributed data and/or dependence among factors and errors is
the Scaled ML (See Ullman 1996 for further discussion).
Whatever function is chosen, the desired result of
the estimation process is to obtain a fitting function
that is close to 0. A fitting function score of 0 implies
that the models estimated covariance matrix and the
original sample covariance matrix are equal.
Model estimation
Indicate software and version
Indicate type of data matrix
analyzed
Indicate estimation method used
Assessing Fit of the Model: AMOS
Output
Divided into
Model, Parameters, and Estimation
Summary
Model Assessment
Model Misspecification
Application 1: Barbara Byrne
p55-93
Parameters and Model Summary
Parameters Summary
Listed are
All the variables in the model according
to their categorization- observed/
unobserved, endogenous/ exogenous
(All factors/ error terms unobserved)
Total number of variables in the model
and for each category.
Summary of the parameters: 32
regression weight, 20 fixed, 12
estimated.
Total of 58 parameters, 38 to be
estimated
Overall Summary of the
Model
Overall Summary
hypothesized model is of recursive type
Sample size is 265
Info to determine identification status
=p(p+1)/2=16(17)/2=136 pieces of
information, 38 to be estimated hence
136-38=98 degrees of freedom
Finally, a summary of the estimation
process chi-square value of 167.153
with p-value of .0000
Limitations of Fit Indices
Values of fit indices indicate only the
average or overall fit of a model. It is thus
possible that some parts of the model may
poorly fit the data.
Because a single index reflects only a
particular aspect of model fit, a favorable
value of that index does not by itself
indicate good fit. That is why model fit is
assessed based on the values of multiple
indices.
Assessment of Model Fit
Identify Offending Estimates
Overall Model Fit
Measurement Model Fit
Structural Model Fit
Identify Offending Estimates
Unacceptable or illogical estimates, such as
Standardized coefficient near or exceeding 1.0
Very large standard error(s)
Constructs negative variance
Remedies
Neg. error variance fix at .005
Assess discriminant validity for highly correlated constructs
Eliminate constructs to stabilize solution
Offending Estimates MUST be eliminated to proceed
Assessment of Model Fit
Examine the parameter estimates
Examine the standard errors and
significance of the parameter
estimates.
Examine the squared multiple
correlation coefficients for the
equations
Examine the fit statistics
Examine the standardized residuals
Examine the modification indices
Model Assessment: Parameter
Estimates
Provide all parameter estimates
(e.g., factor loadings, error
variances, factor variances)
Include the standard errors of the
parameter estimates and its
statistical significance
Need to consider the clinical as well
as the statistical significance of the
parameter estimates
Assessing Parameters
Estimate
Once the model has attained an acceptable fit,
individual estimates of free parameters are
assessed.
Free parameters are compared to a null value, using
a z-distributed statistic.
The z statistic is obtained by dividing the parameter
estimate by the standard error of that estimate.
The ratio of this test must exceed +/-1.96 in order for
the relationship to be significant.
After the individual relationships within the model
are assessed, parameter estimates are
standardized for final model presentation.
When parameter estimates are standardized, they
can be interpreted with reference to other
parameters in the model and relative strength of
pathways within the model can be compared.
Assessing Fit of the Model
A fitting function value of close to 0 is desired for good model
fit. In general, if the ratio between X2 (Chi-Square) and
degrees of freedom is <5, the model is a good fit (Ullman
1996).
To have confidence in the goodness of fit test, a sample size of
100 to 200 is recommended (Hoyle 1995). In general a model
should contain 10 to 20 times as many observations as
variables (Mitchell 1993).
Ullman (1996) discusses a variety of non-X2-distributed fitting
functions, which he calls comparative fit indices. Hoyle (1995)
refers to these as adjunct fit indices. Basically, these
approaches compare the fit of an independence model (a
model which asserts no relationships between variables) to the
fit of the estimated (your) model. The result of this comparison
is usually a number between 0 and 1, with 0.90 or greater
accepted as values that indicate good fit.
Both Hoyle and Ullman suggest use of multiple indices when
determining model fitness.
Measures of Fit
Measures of fit are provided for three
models:
Default Model this is the model that you
specified
Saturated Model This is the most general
model possible. No constraints are placed on the
population moments It is guaranteed to fit any set
of data perfectly.
Independence Model The observed variables
are assumed to be uncorrelated with one
another.
Summary of Models
Overall Measures of Fit
NPAR is the number of parameters being
estimated (q)
CMIN is the minimum value of the
discrepancy function between the sample
covariance matrix and the estimated
covariance matrix.
DF is the number of degrees of freedom
and equals the p-q
p=the number of sample moments
q= the number of parameters estimated
Overall measures of Fit
CMIN is distributed as chi square with
df=p-q
P is the probability of getting as large
a discrepancy with the present sample
CMIN/DF is the ratio of the minimum
discrepancy to degrees of freedom.
Values should be <5 for correct
models.
Cont.Summary of Models
Overall Model Fit
Goodness-of-Fit (GOF) input correspondence of actual or
observed data matrix with that predicted form model (output)
Three basic types of GOF:
Absolute fit - no adjustments
Incremental fit - compare to base model
Parsimonious fit adjust for degree of overfitting,
typically used to compare competing models
Goal: Model parsimony (fewer parameters) for acceptable fit
Evaluating Goodness of Fit:
Definitions of Some Key Statistics
The _
2
Test assesses the overall model in terms of the
discrepancy between the sample and estimated covariance
matrix. If the theory is supported then the _
2
value computed from
the sample must be statistically insignificant (P< .05)
The Goodness-Of-Fit (GFI) is interpreted in a similar manner to R
2
in multiple regression. It is the amount of variances and sample
covariances in the sample covariance matrix that are predicted by
the hypothesised model.(GFI . >9)
The Adjusted Goodness-Of-Fit (AGFI) is analogous to the
Adjusted R
2
in multiple regression analysis. (AGFI > .80)
The Root Mean Square Residual (RMSR) is the square root of the
average of the squared amount by which the sample variances
and covariances differ from their estimates obtained if the
hypothesised model is correct.The larger RMSR the poorer the fit
Performing Model Analysis-I
Evaluation of Overall Model Fit:
Decide on level of significance 10%, 5% or 1%?
Look at Chi-square (_
2
) and GFI absolute fit measures
Look at AGFI, TLI, NFI incremental fit measures
Look at PGFI, PNFI, N_
2
a parsimonious fit measure
Look at the Residual matrix and the Standardised
Residual matrix
Note: In practice values of 0.90 or above for GFI, AGFI
and N_
2
are generally taken to indicate acceptable fit.
Comparisons to a Baseline
Model
NFI is the Normed Fit Index. It
compares the improvement in the
minimum discrepancy for the specified
(default) model to the discrepancy for
the Independence model. A value of
the NFI below 0.90 indicates that the
model can be improved.
Bentler-Bonett Index or Normed Fit
Index (NFI)
Define null model in which all
correlations are zero:
X
2
(Null Model) - X
2
(Proposed Model)
X
2
(Null Model)
Value between .90 and .95 is
acceptable; above .95 is good
A disadvantage of this index is that the
more parameters, the larger the index.
Comparisons to a Baseline Model:
Summary
RFI is the Relative Fit Index This index takes
the degrees of freedom for the two models into
account.
IFI is the incremental fit index. Values close to
1.0 indicate a good fit.
TLI is the Tucker-Lewis Coefficient and also is
known as the Bentler-Bonett non-normed fit
index (NNFI). Values close to 1.0 indicate a
good fit.
CFI is the Comparative Fit Index and also the
Relative Noncentrality Iindex (RNI). Values
close to 1.0 indicate a good fit.
Tucker Lewis Index or Non-
normed Fit Index (NNFI)
Value: X
2
/df(Null Model) - X
2
/df(Proposed Model)
X
2
/df(Null Model)
If the index is greater than one, it is set to1.
Values close to .90 reflects a good model fit.
For a given model, a lower chi-square to df
ratio (as long as it is not less than one)
implies a better fit.
Comparative Fit Index (CFI)
If D= X
2
- df, then:
D(Null Model) - D(Proposed Model)
D(Null Model)
If index > 1, it is set to 1; if index <0, it is set to 0
A lower value for D implies a better fit
If the CFI < 1, then it is always greater than the
TLI
The CFI pays a penalty of one for every
parameter estimated
Root Mean Square Error of
Approximation (RMSEA)
Value: \[(F
0
pop-F
0
est)/n]/df
F
0
is the minimum vale of the discrepancy
function.
If X
2
< df for the model, RMSEA is set to 0
Good models have values of < .05; values
of > .10 indicate a poor fit.
It is a parsimony-adjusted measure.
Amos provides upper and lower limits of a
90% confidence interval for the RMSEA
Absolute Fit Measures
Fit Measure
Likelihood ratio:
Chi-square ()
(Predominant Measure)
Goodness-of-fit:
(GFI)
Root Mean Square:
Residual (RMSR)
Root Mean Square:
Error of Approximation
(RMSEA)
Acceptable Fit
Statistical test of
significance
higher values >0.9
avg residual of input matrix 0
Under .08, best if under .05
Incremental Fit Measures
Fit Measure
Tucker-Lewis
(TLI) or NNFI
Normed Fit Index
(NFI)
Adjusted goodness-of-fit
(AGFI)
Comparative Fit Index
(CFI)
Acceptable Fit
.90 or above
.90 or above
.80 or above
.90 or above
Parsimonious Fit Measures
Fit Measure
Normed Chi-square
( / df)
(Predominant Measure)
Parsimonious GFI (PGFI)
Parsimonious NFI (PNFI)
Akalike information
criteria (AIC)
Acceptable Fit
lower limit: 1
upper limit: 2 or up to 5
higher values
higher values
smaller values
A Model is accepted if the FIVE
following criteria are met;
1. p > 0.05
2. CMIN/DF < 5
3. CFI > 0.9
4. NFI > 0.8
5. Tucker-Lewis > 0.95
6. RMSEA < 0.06
Measurement Model Fit
Three step procedure:
Prove unidimensionality by ensuring significant
structure and no cross-loadings
Evaluate each loadings statistical significance
Assess fit of each construct:
Reliability internal consistency
Variance extracted measures avg. variance
accounted for
Acceptable levels of fit
Reliability : 0.70 or above
Variance extracted: 0.50 and above
Measurement Model Fit Calculations
Reliability measure of internal consistency (not validity),
with common threshold of .70.
Reliability = ( std loading) .
( std loading) + msmt errors
Variance = ( std loading) .
Extracted ( std loading) + msmt errors
Variance Extracted is the degree of shared variance
Model Modification
If the covariance/variance matrix
estimated by the model does not
adequately reproduce the sample
covariance/variance matrix, hypotheses
can be adjusted and the model
retested. To adjust a model, new
pathways are added or original ones
are removed. In other words,
parameters are changed from fixed to
free or from free to fixed. It is important
to remember, as in other statistical
procedures, that adjusting a model after
initial testing increases the chance of
making a Type I error.
Model Modification . Cont.
The common procedures used for model modification are the
Lagrange Multiplier Index (LM) and the Wald test. Both of
these tests report the change in X2 value when pathways are
adjusted. The LM asks whether addition of free parameters
increases model fitness. This test uses the same logic as
forward stepwise regression. The Wald test asks whether
deletion of free parameters increases model fitness. The
Wald test follows the logic of backward stepwise regression.
To adjust for increased type one error rates, Ullman (1996)
recommends using a low probability value (p<0.01) when
adding or removing parameters. Ullman also recommends
cross-validation with other samples. Because the order in
which parameters are freed can affect the choice of remaining
parameters, LM should be applied before the Wald test (i.e.,
add all parameters before beginning to delete them)
(MacCullum 1986, cited in Ullman 1996). Refer to Ullman
(1996) and Hoyle (1995) for further description of these and
other model modification techniques.
SEM and Its Applications
Example 2:
Data with three continuous predictor
variables: education level, a socioeconomic
indicator, and feelings of powerlessness
measured in 1967.
There is one continuous dependent variable,
feelings of powerlessness measured in 1971.
These data are simulated based on the
results reported in a larger study by
Wheaton, Muthn, Alwin, and Summers
(1977).
If you run a multiple regression
analysis in SPSS for Windows using
these variables, you will obtain the
following results;
Now consider the equivalent model fit in
AMOS:
Evaluating Global Model
Fit Using AMOS
Model Re-specification
The Threshold for
Modification Indices
allows you to specify what
level of chi-square change
is required for a path to be
included in the
modification index output.
The default value is 4.00
because it slightly
exceeds the tabled critical
value of a chi-square
distribution with one
degree of freedom: 3.84.
Any additional parameter
estimated by AMOS
should result in an
expected reduction in the
model chi-square of at
least 3.84. The
modification index results
for Example1 can be
obtained)
Example 3
Felson & Bohrnstedts Model (1979)
Perceived academic performance is
modeled as a function of GPA and
perceived attractiveness
Perceived attractiveness, in turn, is
modeled as a function of perceived
academic performance, height,
weight, and the rating of attractiveness
by children from another city.
Model A
GPA
height
weight
rating
attract
error1
error2
academic
.49
.00
-.08
.36
.52
.18
.34
-.27
-.16
.15
-.10 -.08
-.01
P = 0.251 > 0.05
CMIN/DF = 1.381 < 5
CFI = 0.996 > 0.9
NFI = 0.988 > 0.8
RMSEA = 0.043 < 0.06
Conclusion: Model is accepted
Model B: Respecification
GPA
height
weight
rating
attract
error1
error2
academic
.49
.00
-.08
.36
.52
.18
.34
-.27
-.16
.15
-.10 -.08
P = 0.43 > 0.05
CMIN/DF = 0.921 < 5
CFI = 1 > 0.9
NFI = 0.988 > 0.8
RMSEA = 0.000 < 0.06
Conclusion: Model is accepted
Model B is better than Model A
Interpret and Modify Model
Basic Model Interpretation
Diagnosis for Model Re-specification
Process of Model Re-specification
Basic Model Interpretation
Objective = correspondence of proposed model to theory.
Key diagnostic = statistical significance of
estimated parameters.
Standardized versus unstandardized solutions.
Diagnosis for Model Re-specification
Re-specification = improve model fit to best
associate with theory
Diagnostic measures
Estimated parameters:
Std. Residuals = > 2.58
Fixed (non-estimated) parameters:
Modification indices = > 3.84
Expected change parameter
Process of Model Re-specification
Relationships placed in two categories:
Theoretical appropriate for re-specification.
Empirical not open to modification.
Always use theory as guide, not model fit.
Structural Model Fit
Three step procedure:
1. Ensure theoretical support with
statistical significance of path coefficients
(look at directional signals ())
2. Assess coefficient of determination (R) of
each structural equation (size and significance)
3. Ensure that estimated coefficients are not
highly correlated
What to Report in Documenting Results
The covariance matrix between all variables
(constructs) used in the research model
Variable (construct) means and their standard
deviations
Parameter Estimates, that is, the structural
path coefficients and the construct loadings
Overall fit indices. One has to be very
selective in the choice absolute,
parsimonious, comparative fit measures
R-square
Remember
Theory guides SEM in both formulation and
re-specification
SEM is based on two models:
Measurement defines reliability of constructs
Structural estimates causal relationships
Must determine goodness of fit at multiple
levels
Links to other sites on SEM
and Path Analysis:
http://www2.chass.ncsu.edu/garson/pa765/structur.htm
This is an excellent web page. It covers SEM in depth, mostly focusing on goodness of fit tests and
assumptions of SEM. It is very thorough and therefore lengthy. Discussions are related to AMOS, LISREL
and EQS computer applications, especially in terms of capabilities in goodness of fit tests. The discussions
of identification, handling of missing data, and methods for estimating path coefficients are all very helpful.
The end of the page gives information in a Frequently Asked Questions format and also provides links to a
number of SEM packages (under the question heading: Does it matter which statistical package you use for
structural equation modeling?)
http://www2.chass.ncsu.edu/garson/pa765/path.htm
This web page from the above author focuses on path analysis. It is also a very useful site, with discussion
of key concepts and terms, assumptions, and a Frequently Asked Questions section.
http://www.statsoft.com/textbook/stathome.html
This site provides a short explanation of SEM.
http://www.maths.ex.ac.uk/~jph/psy6003/pathanal.html
This site from the University of Exeter provides a brief introduction and explanation of path analysis. It is
part of a set of class notes.
http://www.uic.edu/classes/idsc/ids570/ntspaths.htm
This site from the University of Chicago also provides class notes on path analysis and SEM. It answers a
general list of questions about SEM: why to use it, when to use it and how, etc.
http://www.geocities.com/CollegePark/Campus/caveat.htm
This short paragraph discusses the uncertainty of causal direction in SEM pathways and the problems
caused by post hoc adjustment of models. It provides references to more in depth discussion of these
topics.