Working with Difficult Errors in SEM

Report
Troubleshooting problems with
SEM models that have “Heywood”
cases such as negative variance
parameters and non-positive
definite covariance matrices
Jeremy Yorgason
Brigham Young University
Introduction
• In SEM, it is fairly common to encounter Improper Solutions
• Non-positive definite covariance matrices
• Models with negative variance terms
• Negative PSI matrix
• Correlation or other standardized values > 1
• Model is not identified, you need “x” number of constraints for it to be
identified
• Why is this important?
• Results from models that have this problem cannot be trusted, and
shouldn’t be reported in journal articles.
• Standard errors of estimates may be affected (Chen et al., 2001)
• Error messages are Diagnostic tools
• It’s a good idea to confirm the diagnosis that the computer system is giving
you
• The point is to understand what may be going on with your model/data
• This often requires that you look at all of the output for your model
Goals for this Segment of the Workshop
1. How do I recognize the
problem?
2. How do I fix the problem?
3. Examples
Causes of Improper Solutions in SEM
Causes of Improper Solutions in SEM
1. Specification error in the model
A.Missing a “1” on one of the factor loadings of a latent
variable, or on an error term
B.Correlations of variables or errors from IV to DV of a model
C.Excessive error correlations on indicators of a single latent
variable
D.Very low factor loadings on a latent variable
E.Omitted paths that should be in a model
2. Model under-identified (negative degrees of freedom)
A. V(V+1)/2 minus parms (if estimating means/intercepts use V(V+3)/2)
3. Non-convergence
4. Outliers in the data
5. Too small of sample for the model being estimated
Kline, 2011; Kolenikov & Bollen, 2012; Chen et al., 2001; Newsome, 2012
Causes of Improper Solutions in SEM
6. Missing data
7. “Sampling fluctuations”
8. Two indicator latent variables
A.This includes 2nd order latent variables
9. Non-normally distributed outcome or indicator variables in your
model
A.Categorical
B.Count, zero-inflated, etc.
10. Empirical under-identification
A.“Positive degrees of freedom, but there is insufficient covariance
information in a portion of the model for the computer to generate valid
estimates” (Newsome, 2012)
B.May be caused by some of the above issues
Kline, 2011; Kolenikov & Bollen, 2012; Chen et al., 2001; Newsome, 2012
Signs that there is a problem
Amos:
“XX: Default Model”
“The following variances are negative.”
“This solution is not admissible”
“The model is probably unidentified. In order to achieve
identifiability, it will probably be necessary to impose 1 additional
constraint.”
In place of estimates in the Amos output you see “unidentified”
Signs that there is a problem
Mplus:
THE MODEL ESTIMATION TERMINATED NORMALLY
THE STANDARD ERRORS OF THE MODEL PARAMETER
ESTIMATES MAY NOT BE TRUSTWORTHY FOR SOME
PARAMETERS DUE TO A NON-POSITIVE DEFINITE FIRSTORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE
TO THE STARTING VALUES BUT MAY ALSO BE AN
INDICATION OF MODEL NONIDENTIFICATION. THE
CONDITION NUMBER IS
-0.762D-17. PROBLEM INVOLVING
PARAMETER 59.
MODIFICATION INDICES COULD NOT BE COMPUTED.
THE MODEL MAY NOT BE IDENTIFIED.
More signs that there is a problem
Mplus or other programs:
1. Negative variance estimate (remember Variance =
stddev2)
A.
Find it in your output
2. Correlations above 1 (remember, can’t be larger than 1)
A. Find it in your output
3. Error variance that is really BIG (999 usually indicates a
problem in Mplus, although this is ok if something is
“constrained” to be a certain number)
How can I fix these problems?
1.
Look at a diagram of your model and see if you have miss-specified
your model.
A.
B.
C.
D.
E.
F.
Check your syntax (e.g., look for missing semi-colons in Mplus)
Missing “1” for a factor loading on latent variables
Missing “1” on regression path of error term
Sometimes Amos creates “GHOST” variables. You can’t see them, but they
are there! Sometimes off the screen, sometimes really, really, really, small
Sometimes Amos will “double correlate variables”
Any correlations across IV/DV lines?
1.
Careful, this is something your “modification indices” will suggest to improve
model fit. However, don’t ever add parameters that go against theory
Make sure you have appropriate regression paths in the model (not too
few, in this case)
Make sure your measurement model is appropriate
G.
H.
A.
B.
Factor loadings > .40
Error correlations – start with none, correlation between items is captured in the
latent variables. Typically you’ll use mod indices here
How can I fix these problems?
2. Researchers need to be attentive to model problems
when there are latent variables with only 2 indicators (can
be unstable)
A. Newsom (2012) suggests constraining the two factor loadings to
be equal…
3. Caution is also warranted when estimating “higher order”
latent variables with only two factors, and certain complex
models (e.g., common fate models) that require specific
constraints in order for the model to be identified
How can I fix these problems?
4. Either use a large sample , OR check the sample size
and compare with the number of parameters being
estimated.
• N/q rule (n = sample, q = parameters in the model; Kline, 2011)
• Count variances, covariances, and means OR
• Most programs tell you how many parms are in your model. Amos:
Number of distinct sample moments: 77
Number of distinct parameters to be
44
estimated:
Degrees of freedom (77 - 44): 33
• Quick Check: 10 people in the sample for every observed
(rectangle) variable in the model
How can I fix these problems?
5. If your model looks to be specified correctly, but you still
have a problem with the model, it’s time to start looking at
your data
A.
Run a frequency on all variables in the model, to see if there is
some data entry error or outliers that could be inflating the
variance of one or more variables
1.
2.
3.
Side note: sometimes SEM models have trouble with variables
that have very different (larger or smaller) variance values than
the rest of the variables in your model (e.g., income in dollars)
If this is the case, you will want to rescale or transform these
variables to ensure similar variances
Also, in the transfer of data from one program to another,
sometimes columns of data are shifted or otherwise corrupted
How can I fix these problems?
6. Do you have any categorical or non-normally distributed
dependent variables that are specified as continuous?
A.
B.
C.
Amos doesn’t handle dichotomous, count, or zero inflated
outcomes
Mplus does handle them well, but you have to specify in the
syntax that you are working with such distributions
You may have specified non-normal variable distributions, but
you have small cell sizes (e.g., ordered categorical variable
with only 1 or 2 cases on one end of the distribution)
How can I fix these problems?
• 7. If your model does not “converge” it means that the
program went through X number of iterations, but could
not find a suitable solution. You can increase iterations
from the default number to try to estimate your model. If
this doesn’t work, you probably need to change your
model or you have a data problem.
Atypical Solutions: Start Values and
Iterations
8. A start value is a number assigned to each estimated
parameter when “iterations” begin for a model. Amos and
Mplus automatically create start values for each parameter to
be estimated, yet it is possible to assign start values if the
program assigned ones don’t work. Researchers can provide
start values for a model, which are essentially any known
parameter estimates (e.g., regression weight or coefficient).
You can get these by running simple linear regression with the
variables in your model, and then plug in the coefficient from
the simpler model.
A.
B.
How in the world would I know if I have bad start values???
How would I know what variable to look at that might be nonnormally distributed? Or be categorical and have small cell sizes?
Greek Alphabet and Mplus Output
•
•
•
•
•
•
•
•
•
Nu (Ν/ν)= intercepts or means of observed variables
Lambda (Λ/λ)= Factor Loadings
Theta (Θ/θ)= error variances and covariances
Alpha (Α/α)= means and intercepts of latent variables
Beta (Β/β) and Gamma (Γ/γ) = regression coefficients
Psi (Ψ/ψ)= residual variances and covariances of continuous latent variables
Tau (Τ/τ) = thresholds of categorical observed variables
Delta (Δ/δ) = scaling information for observed dependent variables
Etc. – see Mplus manual
• Ask for Tech1 in the output and then when Mplus says there is a problem with,
for example, parameter #16, go and find that parameter and see which matrix
it is in and then identify the variable and go look at the model/data to see
where the problem is. If no variable is identified, need to go back to Model
Specification.
• CAUTION: Specific Parameter warnings are usually a DECOY! They generally are
simply letting you know the model is not correctly specified, and no matter what you do
to the identified variable it will not make your model work.
Examples: “Message of Death!”
• From a class assignment with a model involving 56 cases.
• Mplus error:
THE MODEL ESTIMATION TERMINATED NORMALLY
THE STANDARD ERRORS OF THE MODEL PARAMETER ESTIMATES MAY NOT BE
TRUSTWORTHY FOR SOME PARAMETERS DUE TO A NON-POSITIVE DEFINITE
FIRST-ORDER DERIVATIVE PRODUCT MATRIX. THIS MAY BE DUE TO THE STARTING
VALUES BUT MAY ALSO BE AN INDICATION OF MODEL NONIDENTIFICATION. THE
CONDITION NUMBER IS 0.383D-13. PROBLEM INVOLVING PARAMETER 31.
THIS IS MOST LIKELY DUE TO HAVING MORE PARAMETERS THAN THE SAMPLE SIZE
IN ONE OF THE GROUPS.
WARNING: THE RESIDUAL COVARIANCE MATRIX (THETA) IN GROUP GRAD IS NOT
POSITIVE DEFINITE. THIS COULD INDICATE A NEGATIVE VARIANCE/RESIDUAL
VARIANCE FOR AN OBSERVED VARIABLE, A CORRELATION GREATER OR EQUAL TO
ONE
BETWEEN TWO OBSERVED VARIABLES, OR A LINEAR DEPENDENCY AMONG MORE
THAN TWO
OBSERVED VARIABLES. CHECK THE RESULTS SECTION FOR MORE INFORMATION.
PROBLEM INVOLVING VARIABLE DE4.
Atypical Solutions: Sampling Fluctuations
• Model is specified correctly
• Don’t have outliers in your data, and you have a large enough sample to
estimate the model at hand
• Model is not identified, although you have positive degrees of freedom
• 9. Possible tests to confirm you have sampling fluctuations and not some
other problem:
• Confidence interval from standard errors includes a zero
• Calculate a “z” by taking the ratio – Estimate: Standard Error, and then compare to a
•
•
•
•
•
•
•
z distribution
Wald test, take ratio – (Estimate:Standard Error)2 then compare to a chi-square
distribution with 1 df
Likelihood ratio test statistic
Lagrangian multiplier (mod indices when var constrained to 0)
Boostrap resampling method (esp. with non-normal data)
Scaled chi-square difference test
Signed root tests
Empirical sandwich estimators
Atypical Solutions: Sampling Fluctuations
• Model is specified correctly
• Don’t have outliers in your data, and you have a large
enough sample to estimate the model at hand
• Model is not identified, although you have positive degrees of
freedom
• Fix the negative variance to 0 or to a small positive number
• This can affect model parameters
Handout
• Chen et al (2001) suggested decision tree
• 1. Is your model identified?
• 2. If so, do you have any negative error variances?
• 3. If so, do you have any outliers that are a problem?
• 4. If not, is the model empirically underidentified?
• 5. If not, do you have sampling fluctuations?
• 6. If so, constrain the negative variance to be 0, a small positive
number, or to be the population variance
• Newsome (2012) prevention tips
• Careful specification
• Use larger samples
• Model factors with 3 or more indicators
• Use reliable measures (high loadings)
• Well conditioned data
Working Example
• See Amos Program
• Depending on time, manipulate an example to show what
errors commonly occur, what the program tells you, and
how to fix the problems
Conclusion
• Either….
• Work with perfect data and perfect models
• OR
• Learn to interpret SEM error messages, and how to fix common
problems
References
• Chen, F., Bollen, K. A., Paxton, P., Curran, P., & Kirby, J.
(2001).Improper solutions in structural equation models:
Causes, consequences, and strategies. Sociological
Methods and Research, 29, 468-508.
• Kline, R. B. (2011). Principles and practices of structural
equation modeling (3rd Ed). New York, NY: Guilford Press.
• Kolenikov, S. & Bollen, K. A. (2012). Testing negative error
variances: Is a Heywood case a symptom of
misspecification? Sociological Methods and Research,
41, 124-167.

similar documents