Panel Data Gothenburg - Sessions 3-4.structual Equation Panel Models

Analysis of Panel Data:
Structural Equation Panel Models
Sessions 3-4: 16 June 2015

Steven Finkel, PhD
Daniel Wallace Professor of Political Science
University of Pittsburgh USA
SEMs: A General Analytic Framework
Also referred to as covariance structure analysis (CSA)
Method for analyzing systems of equations, how a series of variables may relate
together or influence one another, either in a unidirectional causal sense or
in a more complex models of reciprocal causality
Present a system of equations that represents the causal linkages between
variables, so that the model shows all ways that the variables are related. By
estimating the parameters in the model as a whole, as opposed to equation by
equation, we gain information to test more complex models
1960s-90s: one of the dominant methods in all of quantitative social
science. Now other methods compete with SEMs, but still remains a
major part of methodological toolkit for longitudinal and other analyses
Enjoying a recent renaissance
Directed Acyclic Graphs (DAGs) and the approach to causality outlined by Judea
Pearl (2000) and followers
Multilevel Structural Equation Models (MLSEM) in Psychology developed by
MacKinnon, Preacher and others
Analysis of Panel Data, University of Gothenburg, 15-18 June 2015
Example: Miller and Stokes (1956) Model of
Congressional Representation
Illustrates features of SEMs:
Congresspersons System of Equations: here modeling
Perception of three different DVs
District Opinion Exogenous versus Endogenous variables
(as opposed to Independent versus
Dependent variables)
District Roll Call
Direct versus Indirect Causal Effects
Opinion Voting
Recursive versus Non-Recursive Models
Identification of model parameters: is
Congresspersons
there enough information in the
Own Opinion
model to estimate all the coefficients
of interest?

SEMs and Longitudinal Analysis: A Natural Fit
Multiple waves of observations means equations for each DV in each
time period, so can conceive of longitudinal models as systems of
equations over time
Interest in much longitudinal research is in modeling reciprocal effects
between variables, so SEM methods useful
Similarly, interest in much longitudinal research is in modeling
relationships between variables that are purged of measurement error,
and SEM methods are probably the best available for this purpose
Interest in much longitudinal research in direct and indirect effects, so
SEM methods also ideal (e.g. joining groups leads to more civic skills
which then feeds into more participation)
SEMs can be used for Latent Growth Modeling; this represents one
integration of SEMs with other longitudinal modeling methods. More
general Multilevel SEM models for longitudinal data are also possible,
and well cover the basics later in the course
Example: A Simple Longitudinal SEM
1 2
1 2
Y1 Y2 Y3
One variable, say Party Identification , influencing itself over 3 waves (autoregressive
model)
Two equations, one predicting PID in wave 2, one predicting PID in wave 3
Two endogenous variables (PID2 and PID3), one exogeneous variable (PID1)
It is a recursive model, since no causal feedback effects between variables
1 and 2 are the structural effects linking PID in waves 1-2 and 2-3. They are
sometimes in longitudinal models called the stability effects
1 and 2 are the disturbance or error terms, unobserved influences and
idiosyncratic errors that predict PID in waves 2 and 3
If we express all variables as deviations from their respective means, there is no need
for an intercept term in the Y2 or Y3 equations
1 2
1 2
Y1 Y2 Y3
How to estimate the structural parameters in this model?

Run two regressions, one endogenous variable at a time. This is always
possible and will give you the same estimates in simple recursive models as SEM
methods.
Use SEM methods, which uses all the information that is given in the
data (i.e., the variances and covariances of all the observed variables), and
generates estimates of individual structural coefficients as well as tests of
the fit of the model as a whole to the data

SEM estimation procedure (non-technical version)
1. Express the variances and covariances of the observed variables in
terms of the unknown structural effects of the model
2. Manipulate the variances and covariances of the observed variables
to arrive at estimates of the models unknowns. That is, we solve for
the unknowns in terms of the known variances and covariances.
3. If models are overidentified, i.e., more knowns than unknowns,
we can generate predictions about the variances and covariances that
must be true if the model is correct. We then compare the predicted
and actual values of the variances and covariances: if they are the
same, we say the model fits (or is consistent with) the data; if
they are not the same, we reject the model as a whole.
4. Rejected models then are revised; models that fit may or may
not be accepted, however. Many different models may produce a
good fit to the data. Theory is always relevant!!
SEM analysis is also called Covariance Structure Analysis or
CSA: we use the observed variances-covariances to estimate
structural parameters and we use the estimated structural
parameters to generate predicted variances and covariances,
which we then compare to the actual ones in assessing how
well the model reproduces the observed data. Very dialectic!
So with SEMs we move from a world where we care only about

regression coefficents and the fit (R2) of a particular equation to
a world where we care both about regression coefficients
(structural effects), R2, *and* assessing the overall fit of a
given model

How do we do this? A Simple Example
(1) Y2 = 1Y1 + 1
(2) Y3 = 2Y2 + 2
with assumptions E(1 2 ) = E(Y11) = E(Y2 2 ) = 0
What are the knowns in this system? The observed

covariances between PID at time 1 (Y1), PID at time 2 (Y2), and
PID at time 3 (Y3)
If we standardize the variables (i.e. mean of 0, s.d. of 1), we will
have a correlation matrix between the 3 variables with values of
1 on the diagonal, and the correlations between PID1-PID2,
PID1-PID3 and PID2-PID3 in the off-diagonal cells. There
will be 3 distinct pieces of information in the observed
correlation matrix
Step 1: Express Variances-Covariances (Correlations) in
Terms of Unknown Parameters
(1) Y2 = 1Y1 + 1
(2) Y3 = 2Y2 + 2
with assumptions E(1 2 ) = E(Y11) = E(Y2 2 ) = 0
How do we do this? Some basic covariance algebra
Example: Multiply both sides of equation (1) by Y1 and take
expectations (i.e., the long-run average for population)
Y1Y2 = Y1 ( 1Y1 + 1 )
Y1Y2 = 1Y1Y1 + Y11
Cov(Y1Y2 ) = 1Var(Y1Y1 ) + Cov(Y11 )
Cov(Y1Y2 ) = 1
So our model implies that the observed covariance
(correlation here, given standardized variables) between PID
at time 1 and 2 is equal to 1
Our estimate of the models structural parameter 1, then, is
going to be [r(Y1Y2)]
This is the SEM version of the fact that, in a standardized
bivariate regression model, the regression coefficient is
equal to the correlation coefficient between X and Y
Carrying out the covariance algebra in this system of
equations also generates:
Corr(Y2Y3)=2 and Corr(Y1Y3)=12

Observed Versus Implied Correlations
Y1 Y2 Y3
Y1 1
Observed
Correlations Y2 r(Y1Y2 ) 1
Y3 r(Y1Y3 ) r(Y2Y3 ) 1
1
Correlations
1 1
Implied by
the Causal
Model
11 2 1
Step 2: Solve for unknown structural parameters
What are unknowns? 1 and 2
Easy to solve for each:
1 = r(Y1Y2)
2 = r(Y2Y3)
These are two bivariate regression equations, so in

standardized form, the estimate is equal to the
correlation between the IV and the DV
Multivariate and other more complex models, of course,
will have more complex solutions for each unknown, but
the principle will be the same
Step 3: Assess Model Fit
After solving for 1 and 2 , can assess fit for each equation
with tradition measures like R-squared, etc.
But with SEM, can also test --- under some conditions whether
the model as a whole fits the observed data, i.e., do the estimates
produced from the algebraic manipulations from previous slide
successfully reproduce the observed variances and covariances?
In this case, we can test the model. We see that the model makes
the following prediction:
Corr(Y1Y3)=1* 2
So we solve for 1 and 2 and then generate a prediction for
Corr(Y1Y3). If it is close (given sampling error), then the model is
consistent with the data; if not, we reject the model and say that
it needs to be modified.

Overidentified Models
Assessing model fit is only possible when models are overidentified,
when there are more knowns than unknowns. In our case we
have 2 unknowns (1 and 2) and 3 correlations, so the model is
overidentified. Technically we say that the model has 1 degree of
freedom.
When models are overidentified, they will not necessarily
reproduce the observed data exactly, thus giving us the degree(s)
of freedom to test the model.
When models are just-identified, there are an equal number of
knowns as unknowns, and the model estimates will reproduce he
observed data (correlations) exactly. Also known as a saturated
model

Another way to look at it: there are actually two ways of
arriving at 2
One is our original estimate: r(Y2Y3)
Another is (following covariance algebra multiplying equation
for Y3 by Y1) :
r(Y1Y3)/ r(Y1Y2)
So the test of whether the two estimates are the same,
given sampling error, is whether:
r(Y2Y3)= r(Y1Y3)/ r(Y1Y2)
If so, the model is consistent with the data. If not, reject the
model as specified and need to modify or abandon it

Example: National Election Panel Study,
2000-2002-2004
Stability of Party Identification
1
Observed .872 1
.856 .884 1
1
Implied 1 1
11 2 1
Estimates and Model Fit
1 = r(Y1, Y2) = .872
2 = r(Y2,Y3) = .884
Does the model fit the data?

Predicted r(Y1 Y3)=12=.872*.884=.771
But observed r(Y1 Y3) =.856, so our model is
wrong (without considering sampling error)
Alternatively:
2= r(Y1 Y3) / r(Y1 Y2) =.856/.872=.98
A much different from the earlier estimate of .884
What is wrong with the model? No direct effect from
PID2000 PID2004 (probably, but could be other omissions)

Alternative Model: Include direct effect of PID2000 on PID2004
1 2
1 2
Y2 Y3
Y1
3
Observed Correlations Model Implied Correlations
Y1 Y2 Y3 Y1 Y2 Y3
Y1 1 Y1 1
Y2 r(Y1Y2 ) 1 Y2 1 1
Y3 r(Y1Y3 ) r(Y2Y3 ) 1 Y3 1 2 + 3 2 + 13 1

This model has three unknowns (1, 2 , and 3) and three
knowns
It is saturated or just identified, with 0 degrees of freedom. In
recursive models it means that no structural effects are left out of
the model
As noted earlier, you cant test these models in terms of fit
because the model will necessarily reproduce the observed
variances and covariances exactly
You can see that we are going to manipulate all of the
correlations to obtain estimates of the three unknowns, and so
we have no excess correlations with which we can test the fit
of the model (like we did in the last model with Corr(Y1Y3 )
We can say that the restriction that 3=0 in the previous model
is what allowed us to test it it is the overidentifying
restriction of that model

Maximum Likelihood Estimation of SEM Models
Basic idea: Find the parameters that generate an implied population

variance-covariance matrix ( ) of the observed variables that comes
as close as possible to the actual variance-covariance matrix S. Those
parameters then maximized the likelihood of having observed the
variance covariance matrix we did observe
Test significance of individual coefficient estimates
Use summary statistics to evaluate the fit of the model as a whole:
Modify and compare alternative models, especially models nested
within one another, i.e., with same variables and structure aside from
one or more constraints on parameter values in a reduced versus a
full model

How to Estimate SEM Models via ML?
Two matrices:
S = Observed Variances and Covariances in our sample, which was
randomly drawn from a population with variance-covariances
= Implied Variance-Covariance based on unknown model parameters
(in our case and the variances-covariances of the errors)
THE METHOD OF MAXIMUM LIKELIHOOD FINDS THE
VALUES OF THE UNKNOWN MODEL PARAMETERS
THAT, TAKEN TOGETHER, MINIMIZE THE
DIFFERENCE BETWEEN THE VALUES OF S AND

Technically:
ML estimates are those that minimize a fit function (call it F) that
and S as:
expresses the deviation between
^

S
F = log + tr ^ log S t

where the first and third terms are the determinants of and S, respectively,
tr is the trace of a matrix, and t is the total number of variables in the model.
ML estimates are produced via iterative processes, where initial values are
adjusted, recalibrated, etc. until no better estimates can be produced.
If you are interested in learning about different algorithms used in ML
estimation of this sort, see Kenneth Bollen, Structural Equation Models
for Latent Variables (John Wiley and Sons, 1989), pp.131-144.
When and S are equal, F will be 0, so the goal is for F to be as small
as possible. It will be 0 only when the model perfectly reproduces S,
e.g. with a saturated model.

Testing Model Fit
H0: The covariance matrix in the population is equal to , the covariance
matrix implied by our model. (Informally: our model fits the data, taking
sampling error into account)
HA: The covariance matrix in the population is equal to any arbitrary
matrix, and if it is arbitrary, we will pick S when maximizing the likelihood
function. (Informally: no model at all is needed to estimate, only our
sample variances and covariances).
So under each of these two hypotheses, we can generate a (log) Likelihood
function that summarizes the overall likelihood that the specified model
holds in the population. We can then test H0 by comparing the ratio of the
two (log) Likelihoods.
The ratio of the two log Likelihoods (the Likelihood Ratio Chi-Square,) is
equal to (N-1)*F, and is distributed as a 2 statistics with degrees of freedom
equal to ((t(t+1)/2) k), where t is the number of variables, and k is the
number of unknown parameters in the model. Here we have 6 knowns and
5 unknowns, so df=1.
This means that the Likelihood Ratio chi-square (call it L2) can be
used as a statistical test of the null hypothesis that the specified model
holds in the population.
Note that this is fundamentally different than other null hypotheses
you have encountered, where the null represents NO relationship or
no effect from one variable to another. Here our null is that the
model fits the data, and we therefore DO NOT want to reject it. So
large Likelihood ratio chi-square values means we reject the null and
our model does not fit the data, or is inconsistent with the data.
Our example: (N-1)*F, the Likelihood Ratio Chi-square. Here it is 738*.17 =
110.7, close enough to the 109.8 from STATA output. With 1 degree of freedom,
the statistical significance of this figure is extremely high, that is, it was very unlikely
to have come about by chance sampling error. So we can say that the deviations
between and are statistically significant, that is, there is a true difference between
our and the true population covariance matrix .
This means our model is inconsistent with the data.

Testing Alternative Models
The same logic can also be used to test the statistical significance of any improvements
in model fit associated with relaxing model constraints, that is, in comparing the fit of
two different models.
Suppose we relax the constraint that 3 = 0. Whenever one model can be arrived
at by imposing one or more constraints on another model, we say that the model is
nested in the other model. In such cases, the difference in 2 between the two
models itself follows as 2 distribution, and the improvement in fit can be tested
according to whether the difference in 2 is statistically significant, given the
difference in degrees of freedom (or the number of imposed constraints) between
the two models.
In this case:
Constrained Model (where 3 = 0: 2 = 109.8 w/ 1 df
Unconstrained Model (where 3 is a free parameter): 2 = 0 w/ 0 df
2 Difference (Constrained Unconstrained) = 109.8 df Difference=1
2 Difference is statistically significant and therefore the Unconstrained model
represents a statistically significant improvement over the Constrained model in terms of
reproducing S (or ).
Notes on Model Testing and
Assessing Goodness of Fit
1. In this case the unconstrained model is a saturated model but this need not
be the case. An unconstrained model is any model that relaxes one or more
constraints on model parameters, and the difference in 2 between the two
models provides a test of the significance of the improvement in fit from
relaxing the constraints.
2. It is also the case that this is the only kind of test that you can do with
saturated models, that is, compare them to other models in terms of fit. You
cannot conclude anything about a saturated model in itself because the fit is by
definition perfect.
3. An insignificant model 2 or L2 does not mean that your model is correct
or the true model. It is just consistent with the data. There could be
many models that reproduce the observed data as well as yours. So a
significant model 2 or L2 means that the model is inconsistent with
the data and can be rejected, but an insignificant 2 or L2 does not
mean the model is accepted as the truth.

4. Since the Likelihood Ratio 2 (L2) depends on N and F, you can always increase
the size of 2 by increasing the number of cases. This means that the same
degree of fit will look worse with a larger N, a somewhat perverse situation in
social science, as we usually want to maximize the number of observations.
5. Hence methodologists have come up with various ways of correcting L2.
Sometimes the correction is done by dividing L2 by the degrees of freedom in the
model, but no one knows how to interpret L2/df with any specificity. It is said
that values approaching 2-3 are beginning to be reasonable.
One measure along these lines that is very popular in the SEM literature is the
Root Mean Square Error of Approximation (RMSEA),:
m2 df m
RMSEA =
(N 1)df m
which in this case is: square root of 108.76/737=.384
It is desirable for this value NOT TO EXCEED .10 for models to be acceptable!!!

6. Other measures of goodness-of-fit involve comparing the
model chi-square in various ways to a very crude baseline
model which has no parameters whatsoever and no covariances
between the observed variables it just has variances of
variables that are completely unrelated in the population. (This
is about as baseline as you can get!
Note the baseline model has 3 df since there are 3 observed

variables ( i.e., 3 variances are estimated and that is it!)

We can calculate the extent to which our model improved on the
baseline model via a proportional reduction in the chi-square value, or
the Comparative Fit Index (CFI):
2 2

CFI = b m
2
b
So the CFI is (2288.51-109.77)/2288.51=.952

Tucker and Lewis (1973) suggest correcting the CFI to take into
account the relative degrees of freedom in the two models, so that
near-saturated models are penalized to a greater extent:
b2 m2
dfb dfb
TLI =
b2
dfb 1
which equals (762.8-109.76)/761.8=.857

CFI and TLI should be greater than .90 and close to .95 if possible!!!!
7. Last measures of Goodness of Fit
A. The Average residual, or square root of the total squared
(standardized) deviations between the observed and predicted
covariances. This is the Standardized Root Mean Square
Residual (SRMSR) in STATA, and should not be above .08 (some
say .05) for model acceptability. Here we obtain .028, so not bad.
B. Information-based measures. These are the only measures that do
not depend on models being nested in one another. Akaikes
Information Criteria (AIC) is:
AIC = 2ln Lm + 2 * k
where L is the models log-likelihood(NOT CHI-SQUARE) and k is
the number of estimated parameters. Smaller numbers are better, so
the 2*k term is a penalty for the expression as more parameters are
estimated. Used very often to compare models with different
variables, structures, etc., and we will see this in later sections of
the course as well.
Here AIC=-2(-3835.1)+2*5=7682.2
All STATA Goodness of Fit measures from estat gof, sta(all)

Cross-Lagged and Reciprocal Effects SEM Models
I: The Two-Wave Cross-Lagged Panel Model
Y1 Y2
1
pid2000 pid2002 1
2
rX1Y1
4
app2000
3 app2002 2
X1 3 X2
SEM Diagram in STATA with

Party ID and Presidential
Approval, 2000-2004

Features of the Cross-Lagged Model
A Dynamic Model of Change in Each Endogenous Variable
(a) Y2 = 1Y1 + 2 X 1 + 1 "static score"

(b) Y2 Y1 = Y = ( 1 1)Y1 + 2 X 1 + 1 "change score"
Goal: Compare relative direction and magnitude of cross-lagged effects to

give evidence regarding causal direction and strength
Each equation by itself is a typical longitudinal regression model with a
lagged DV (lagged endogenous variable) and a lagged IV
The estimated effect of X1 is the same in (a) or (b). The latter is sometimes
called the conditional change model the effect of X1 on changes in some
variable, conditional on the initial level of that variable. With no lagged DV
(here no Y1), we would have an unconditional change model

Why Include the Lagged DV?
Substantive Reasons:
Yt-1 may cause Yt (budgets, wealth, e.g.) or Y may have inherent persistence (eg.
states such as marriage, employment, or attitudes
Statistical Reasons
Regression to the Mean: random positive (negative) errors from t1 will recede
(increase) at t2, leading to a negative relationship between Yt-1 and change
If endogenous selection into a treatment based on Yt-1, need to control
SEM solution to general temporal dependence: what causes unit i to be high at
one period may cause it to be high at next, over and above the other factors in
the model. Clustered observations, and in SEM, lagged DV is common solution
Modeling Long-Term System Dynamics and Cumulative Effects
Allows long-run cumulative effects of the IV, with the magnitude of the long-run
effect dependent on the drag to the system exerted by the lagged DV
Equilibrium level of Y given by h* = ( 2 ) X1
1 1
So 2 is the short-term effect, and ( ) is the long-run effect (the Long-
2
Term Multiplier) 1 1
Systems stabilize at equilibrium points only if 1<1

Causal Lags and Synchronous Effects
Cross-lagged model assumes no synchronous or contemporaneous
effect at time t.
Does time period separating waves of measurement correspond to the length of
time for causal effect to take place?
If no, may need to add an additional effect from X2 to Y2 , or have a model that
only has the contemporaneous effect and no lagged effect. Need to think about
how the causal effect operates (this will become even more urgent when we
include the U term for unobserved heterogeneity into the models later)
With a unidirectional single equation model, can add
contemporaneous effect with no problem. Model is still recursive.
Y2=5 X 2 + 1Y1 + 2 X 1 + 1
And since the following is true,

X 2 =X1 +X
can express this as:
(a) Y2 = 1Y1 + ( 5 + 2 ) X 1 + 2 X + 1
Or, in terms of X and Y :
(b) Y = ( 1 1)Y1 + ( 5 + 2 ) X 1 + 2 X + 1
In Time Series language:
Model a: Autoregressive Distributed Lag Model (ADL, 1, 1)
Model b: Error Correction Model (ECM)
BUT in a multiple causal system, including both synchronous and
lagged effects in both directions makes the model non-recursive, and
underidentified. Need more information, as we will see !!!

Estimating the Two-Wave Cross-Lagged Model
It is a recursive model with no reciprocal effects specified at the
same point in time. Therefore, the model is either identified or
overidentified.
Which is it?
Number of knowns: 4 variables=(4*5)/2= 10 covariances
Number of unknowns: 9 (4 , 2 variances of structural disturbances,
plus necessary to estimate the three variances and covariances of the
observed wave 1 exogenous variables)
Therefore, overidentified with 1 degree of freedom
What is left out of this model? What is the overidentifying restriction?
Answer: the covariation between wave 2 disturbances call this 21=0
So the model chi-square tests whether this restriction is correct, i.e.,
whether the cross-lagged and stability effects in the model completely
account for the cross-sectional covariation between Y2 and X2.
Can saturate the model by relaxing this restriction and estimating 21=0.
Smaller values for chi-square are better (why?). The other estimated
coefficients in this revised model will be exactly the same.

STATA Results (Standardized)
>
Likelihood ratio
chi2_ms(1) 81.500 model vs. saturated
p > chi2 0.000
chi2_bs(5) 1419.306 baseline vs. saturated
p > chi2 0.000
1.4 .79 >

pid2000 pid2002 1 .23
1 .11 Population error
RMSEA 0.330 Root mean squared error of approximati
> on
90% CI, lower bound 0.272
.62
.43 upper bound 0.393
.13
pclose 0.000 Probability RMSEA <= 0.05
>
1.5 Information criteria
app2000 .16
app2002 2 .71 AIC 10452.701 Akaike's information criterion
1 1.6 BIC 10512.552 Bayesian information criterion
>
Baseline comparison
CFI 0.943 Comparative fit index
TLI 0.715 Tucker-Lewis index
>
Size of residuals
Modification indices SRMR 0.035 Standardized root mean squared residua
> l
Standard CD 0.791 Coefficient of determination
MI df P>MI EPC EPC
Structural
>
pid2002 <-
app2002 77.161 1 0.00 .2556865 .1830259
app2002 <-
pid2002 77.161 1 0.00 .4089136 .571251
cov(e.pid2002,e.app2002) 77.161 1 0.00 .4519648 .3233477
EPC = expected parameter change

Saturated Version of Two Wave Model
>
Likelihood ratio
p > chi2 .
p > chi2 0.000
1.4 >
.79
pid2000 pid2002 1 .23 Population error
1 .11 RMSEA 0.000 Root mean squared error of approximati
> on
.62 upper bound 0.000
.43 .32
.13 pclose 1.000 Probability RMSEA <= 0.05
>
Information criteria
1.5 AIC 10373.201 Akaike's information criterion
app2000 .16
app2002 2 .71 BIC 10437.656 Bayesian information criterion
1 1.6
>
Baseline comparison
>
Size of residuals
SRMR 0.000 Standardized root mean squared residua
> l
CD 0.773 Coefficient of determination
>

Extension to Three and More Waves
1 2
pid2000 pid2002 pid2004
app2000 app2002 app2004
3 4
Flexibility in model and lag specification increases substantially

with 3 or more waves. Many more additional knowns in the
covariance matrix, yields more degrees of freedom to test
restrictions in the model and compare alternative nested models
Testing Alternative Multi-wave Models
Imposing Equality of Causal Effects is Straightforward
Equal stabilities across waves
Equal cross-lagged effects across waves
Equal error covariances in waves 2 and 3
Note: error covariance in wave 1 cannot be equated with these two.
(Why?)
Test these equality constraints via difference in chi-square
statistics, as each models with the respective constraint is
nested within the model without the respective constraint
Add exogenous background variables
Relax constraints regarding error covariances for same
variable over time (i.e. autocorrelated disturbances)

1 2 3
pid2000 pid2002 pid2004

No Wave 1-Wave 3 PID
Error Term Covariance
educ
app2000 app2002 app2004
4
5 6
1 2 3
Wave 1-Wave 3 PID Error pid2000 pid2002 pid2004
Term Covariance Added educ
app2000 app2002 app2004
4
5 6

Problems/Omissions in the
Cross-Lagged Models So Far
No synchronous causal effects
No measurement error
No controls for stable unmeasured variables that were
focus of Day 1

II. Non-Recursive SEM Models
Involve reciprocal causality between variables at a given point in

time (as opposed to cross-lagged effects)
Note difference between a model with a synchronous causal effect from
one variable to another (which would not necessarily be non-recursive),
and a model with reciprocal synchronous effects, which would necessarily
be non-recursive
When appropriate?
When theory says the causal lag is short, relative to time between waves
When cross-lagged effects are weak or insignificant, with relatively large
covariances between the two variables at a given point in time

Reciprocal Effects or Simultaneous Causality
(5)
Y1 = 1 X 1 + 2
X 1 = 2Y1 + 1
X 1 = 2 ( 1 X 1 + 2 ) + 1
X 1 = 2 1 X 1 + 2 2 + 1
So X1 is related to 2, hence endogenous! Same with Y1 and 1 in

equation predicting X
Further problem: Underidentification!! 4 unknowns, 3 knowns!!
Identification
Model Identification: Use the counting rule.
Number of knowns: t(t+1)/2, where t is the number of observed vars
Number of unknowns: k parameters to be estimated
Model as a whole is identified if t(t+1)/2-k0
But still possible for individual equations not to be identified
A 99.9% effective rule for identifying equations within non-recursive models
is the order condition:
If equation involves p endogenous variables, then there must be at least (p-1)
excluded exogenous (Z) variables for the equation to be identified. Another
way to put it: for each endogenous variable that is included as a predictor,
there must be at least one excluded exogenous variable for the equation to be
identified.
That is, must have a Z that does *not* have an effect on the endogenous
variable in question for every endogenous variable that does. This represents
an exclusion restriction that allow identification of the given equation (i.e.
the X is excluded from the X or Y equation in question).
These Z variables will be the instrumental variables for the causal system
In our reciprocal effects case: to identify each equation, we
therefore need a variable that is a) exogenous, and hence
unrelated to all equations error terms; b) directly affects one
of the endogenous variables in the reciprocal effects causal
system (and the stronger, the better); but c) is excluded, i.e.,
does not affect the other endogenous variable in question.
THESE VARIABLES ARE NOTORIOUSLY
DIFFICULT TO FIND!!!!
IMAGINE TWO VARIABLES THAT ARE
RECIPROCALLY RELATED. IS IT EASY TO ALSO
IMAGINE AN EXOGENOUS VARIABLE THAT
AFFECTS ONE AND ONLY ONE OF THOSE TWO
VARIABLES??

This is one of the main advantages of panel analysis: the lag values
of the variables can, under certain conditions, serve as instruments
for reciprocally-related variables.
If one can assume no cross-lagged effects, and if one can assume
exogeneity for the lagged values, then panels provide ready-made
instruments for causal systems that contain contemporaneous reciprocal
effects.
But if there are cross-lagged effects, or if there are correlations between
the disturbances, these assumptions will not hold. In some cases,
moreover, it will not be possible to test all of these assumptions due to
the lack of information in the causal system.
Later will see about using panel data and the lagged structure of the data
to help with other kinds of endogeneity problems, not only those due
to reciprocal causation.
General Moral: we always estimate causal effects within the

context of models whose assumptions need to be justified,
and which may ultimately not hold

The Two-Wave Synchronous Effects Model
pid2000 pid2002
app2000 app2002
Y2 = 3 X 2 + 1Y1 + 1
X 2 = 4Y1 + 2 X 1 + 2
ML Estimation of the Model
Each of the equations is just-identified: one included
endogenous variable and one excluded exogenous variable.
X1 is an instrument for X2 in the Y2 equation
Y1 is an instrument for Y2 in the X2 equation.
We can arrive at the coefficients through algebraic manipulation
multiply each endogenous equation by each of the wave 1 variables
(X1,Y1) and solve for the 4 structural effects in terms of the observed
covariances. This is tedious algebra but follows the same logic as
what we have been doing so far
ML estimation also follows the same logic as before.
10 observed variances-covariances, 9 unknowns (4 structural effects,
2 disturbances, 3 variances-covariances of wave 1 variables)
Model as a whole has one over-identifying restriction (either one of
the cross-lagged effects is NOT zero, or the disturbance covariance
between the is NOT zero).
The model 2 provides a test of the restriction.

Extensions
Three wave synchronous effects model provides additional
information from third wave, and ability to constrain effects to
be equal across waves to gain additional degrees of freedom.
Two wave combined synchronous and cross-lagged model is
not identified without including more instruments, but since
only two waves exist in the panel, such instruments cannot
come from lagged values of the variables. Hence the
researcher is in the same difficult position as the cross-
sectional analyst in looking for appropriate instrumental
variables in a two-wave model with both lagged and
synchronous effects.

Three wave combined synchronous and cross-lagged model
may be identified in several ways:
Use the wave 1 variables to identify the wave 3 equations. This
is rarely implemented because one cannot say anything about
the wave 2 estimates
Impose equality constraints on the coefficients across waves, so
that the cross-lagged effects are equal from waves 1-2 and 2-3,
and the synchronous effects are also equal between wave 2 and
wave 3.
In the absence of strong theoretical reasons for expecting
different lag lengths, the analyst will often estimate cross-lagged
only, synchronous only, and various combined models, often
with equality constraints as well. Decisions about the likely lag
length and the direction and magnitude of the causal effects are
made after inspecting the overall pattern of results from these
models. In this case, for example, it looks pretty clear that PID
has a synchronous effect on APP, and APP has a lagged effect
on PID.

1 2
pid2000 pid2002 pid2004 Stata estimation of Synch and C-L effect

e
d a b
e
d
a b
model, with effects equated across waves
app2000 app2002 app2004
3 4
Likelihood ratio
chi2_ms(6) 134.899 model vs. saturated 1 .23 2 .21
p > chi2 0.000
chi2_bs(14) 3458.563 baseline vs. saturated 1.4
.82
p > chi2 0.000 pid2000 .8 pid2002 pid2004
1 .11 -.093
Population error .1
.62 .12
RMSEA 0.171 Root mean squared error of approximation -.0092 .52 .0089 -.008 .48 .0098
upper bound 0.196 1.5
.1
pclose 0.000 Probability RMSEA <= 0.05 app2000 app2002 .43 app2004
1 1.5 .16
AIC 14813.960 Akaike's information criterion 3 .65 4 .36
BIC 14910.643 Bayesian information criterion
Baseline comparison
Size of residuals
SRMR 0.029 Standardized root mean squared residual

. estat mindices
Modification indices
Modification indices shows some unexplained
covariation between PID2000 and PID2004
Standard
MI df P>MI EPC EPC
(Presidential Year Effect). Options: add
Structural
pid2002 <- either structural disturbance covariance, or a
pid2004 102.723 1 0.00 -.398454 -.4252892
app2004 15.564 1 0.00 -.1289866 -.1068166
direct causal effect from PID2000->PID2004
pid2004 <-
pid2000 105.332 1 0.00 .3849349 .3561638
app2000 24.988 1 0.00 .1498549 .1073146
app2004 <-
app2000 6.171 1 0.01 .0705983 .0651619
cov(e.pid2002,e.pid2004) 113.023 1 0.00 -.502823 -.4459717

So: final model (?) below:
EPC = expected parameter change
Likelihood ratio
1 .23
p > chi2 0.001
2 .18
p > chi2 0.000
pid2002
.13 .52
Population error 1.4 .8
RMSEA 0.067 Root mean squared error of approximation pid2000 pid2004
1 .36 -.14
upper bound 0.097 .12
.57 -.011 -.041
.11
pclose 0.141 Probability RMSEA <= 0.05 .62
-.048 .52 -.012
AIC 14702.656 Akaike's information criterion 1.5
.1
BIC 14803.943 Bayesian information criterion app2000 app2002 .42 app2004
1 1.5 .16
Baseline comparison
CFI 0.995 Comparative fit index 3 .65 4 .36
Size of residuals
Excellent GOF measures, though

chi-square is still significant

III. Measurement Error Models
The assumption in OLS regression that variables are

perfectly measured is obviously violated in most
research situations, but it is very difficult to
overcome the problem in the cross-sectional context.
With panel data, the information on the same
variables over time allows much more flexibility in
estimating true effects once measurement error is
taken into account, and in estimating the
measurement properties of the indicators as well. In
the SEM framework, all of these effects can be
estimated simultaneously

Measurement Error: The Basics
Indicators of variables may contain error, in that the value that is
assigned to a given unit is not the true value for that variable for
that unit. Errors in variables may be:
Systematic, in which case we may say that the observed indicator is always off
from the true value in one direction or the other, or that some other variable is
also systematically influencing indicator, aside from the true variable of interest.
In that case, we say the measure is not a valid indicator.
Random, in which case the observed indicator is sometimes higher or lower
than the true value depending on random factors in the measurement process,
such as (among other things):
poor record keeping
individual coder decisions (e.g. the people at Freedom House deciding on a 2 versus a 3 for
some countrys civil liberties index).
ambiguous questions in surveys
mood or other transient factors in the interview or observation process
scaling of variables (e.g. where does 3.55 attitude go on a 1-2-3-4-5 scale? Most of the time
to 4, but some of the time no doubt to 3).

Random errors lead indicators to be unreliable
ALL SOCIAL SCIENCE MEASURES ARE

UNRELIABLE TO SOME EXTENT!!!!
SOME ARE ALSO INVALID, THOUGH THIS
PROBLEM IS MUCH MORE DIFFICULT TO
DETECT AND CORRECT.
SEM methods useful for correcting for reliability problems only!

The Problem of Measurement Error for
OLS Regression
Assume the following cross-sectional model: an indicator y* is a
function of some latent construct Y and some random
measurement error wi:
(1) yi* = Yi + wi
Note regression coefficient of 1 for effect of Y on y*. This is a
convenient assumption that simply means that the two variables,
the observed indicator y* and the latent unobserved variable Y, are
measured on the same scale (dollars, miles, number of conflicts,
etc.). Since we do not observe Y or other latent variables, we
always need to set their scale to something.
(This is implicit in the regression coefficient of 1 for any
equations error term as well. The error term is unobserved and
we conveniently set its scale to be the same as the endogenous
variable in its equation.)
Equation (1) above is likely to part of an overall causal system that has a
measurement portion as well as a structural portion where the relationships among
the Y variables are of theoretical interest as well. So assume that Y is caused by
one exogenous X variable, itself with an imperfect x indicator:
(a) Yi = 1 X i + i
(2) (b) yi* = Yi + wi
(c) xi* = X i + vi
Structural part of model is (2a); measurement part is (2b)-(2c)
Assumptions:
E(w)=E(v)=E()=0 (all errors have a mean of 0)
E(X)=E(Xv), E(Yw)=0 (all errors are uncorrelated with their own
equations independent variables)
E(v)=E(v)=E(vw)=0 (all errors are uncorrelated with each other)
, v and w are normally distributed (as is the standard OLS assumption)

Measurement Error in X
Y is related to True X, but we use fallible xi*
that contains random error vi
Yi = 1 X i + i
(3) and x *i = X i + vi
then Yi = 1 (x vi ) + i *
i
and Yi = 1x *i + ( i 1vi )
xi* is related to the error term (since vi and xi* are related). So
fallible xi* is endogenous and we get inconsistent estimates of the
causal effect 1

Measurement Error, Reliablity and
the Attenuation of OLS Estimates
Another way to look at this: the OLS estimate of 1 with the
fallible measure x* is:
= Cov(Yi xi *)
OLS
Var(xi *)
With a little substitution from 6a, we can arrive at:

Cov( TRUE X i + , xi* ) Cov( TRUE X i xi* ) Cov(xi* X i )
OLS = *
= *
= TRUE
Var(xi ) Var(xi ) Var(xi* )
Multiplying 2c through by X and taking expectations shows that

Cov(xi*, X)=Var(X), so:
Var( X i )
(4) OLS = TRUE
Var(xi* )

Conclusion: Unless there is no measurement error in x, the
(bivariate) OLS estimate of will be less than the true value,
and will be attenuated by the factor Var( X i )
Var(xi* )
We call this factor, which is the ratio of true score variance to
observed score variance, the reliability of x (denoted as xx). It is
the proportion of the observed variance in x that is composed of the
latent true score and not the measurement error vi . So the OLS
bivariate equals the true multiplied by the reliability of x
Higher reliability means that the observed score is closely related to

the true score and hence the attenuation of the OLS regression
coefficient will be small; lower reliability means greater random noise
in the indicator and consequently greater attenuation of the OLS
regression coefficient in the bivariate case

Notes on Reliability
The direction of bias due to measurement error in explanatory
variables is always downward in the bivariate case; in the
multivariate case it may be downwards or upwards, depending on
the amount of measurement error in particular variables and their
intercorrelations
We can also take equation 2c, square both sides and take
expectations to yield:
Var(xi* ) = Var( X i ) +Var(vi )
which expresses the variance in a fallible indicator as composed of

two parts: the true score variance and the error variance. So the
reliability of x is the proportion of its variance being true score
variance it is akin to R2 in that we can say that the higher the
reliability, the lower the error variance in an indicator and thus the
greater amount of variance explained by the latent true score X
Measurement Error in Panel Models
For panel analyses, measurement error is especially problematic.
It turns out that measurement error in the lagged endogenous
variable wreaks havoc with the estimation of a variables stability,
and hence throws off inferences in nearly all of the models that
we have examined thus far
The ability to model the substantive effect of lagged y will be
compromised, along with the ability to use lagged y as a control
for regression to the mean effects and for the estimation of the
models causal dynamics and long-range effects of X
Thus --- somewhat ironically --- correcting for measurement error
is one of the key advantages of panel analyses, but it is also with
panel data that measurement error corrections are sorely needed

Example: Two Wave Y-Only Model with
Measurement Error
1 2
1
Y1 Y2
y1* y2*
w1 w2
MEASUREMENT MODEL FIGURE 1
NOTE: SEM convention to use circles to represent latent variables,

squares to represent observed variables. Also using wave 1 as endogenous for ease
of presentation; same concepts apply if wave 1 indicator is x* and latent variable is X

We can write the structural portion of the model in equation form as:
(5a) Y1 = 1
(5b) Y2 = 1Y1 + 2
And the measurement portion as:
(6a) y1* = Y1 + w1
(6b) y2* = Y2 + w2
Subtract equation 6(a) from 6(b) to arrive at an expression for the change in y:
y* = Y + (w2 w1 )
Multiply equation by y*1 in the first step, substitute (Y1+w1) for y*1 in one term on
the right side of the equation in the second step, and take expectations to yield:
(7) Cov(y*, y ) = Cov(Y ,Y1 ) Var(w1 )

*
1
THIS IS A VERY IMPORTANT RESULT!!!

It says that the covariance between the initial value and the change in a fallible
indicator over time is equal to the true covariance between the latent initial
true score and the change in the latent true score over time MINUS the amount
of measurement error in the initial (time 1) indicator.
So the more measurement error in y1, the more we overestimate the amount of
negative covariation between the variables initial level and change if we dont
take the measurement error into account. That means that, with measurement
error in the indicators, we think that there is more regression to the mean than
there really is and more speed toward equilibrium than there really is.
In terms of the level of the dependent variable, we think there is less stability
from one time point to the next when we analyze the fallible indicators than
there really is in the true score latent variables. And to the extent that
exogenous variables X are related to the initial level of y,* we will consequently
wrongly estimate their impact on y2* as well.
Big problems, right?

Correcting for Measurement Error in SEM Panel
Models: Single Indicators, Multiple Waves
Consider again the causal system depicted in Figure 1 (Equations 5 and 6):
The structural portion:
(5a) Y1 = 1
(5b) Y2 = 1Y1 + 2
The measurement portion:
(6a) y1* = Y1 + w1
(6b) y2* = Y2 + w2
Are the model parameters identified?
How many unknowns? Five:
One structural effect 1
Two variances of the disturbance terms 1 and 2
Two measurement error variances of the w1 and w2
How many knowns? Three: variance of y* at time 1, the variance of y* at time 2,
and their covariance.
So the two wave, single indicator model is underidentified!! We need either more
waves or more indicators, or both!
Three Wave, Single Indicator Model
1 2 3
1 2
Y1 Y2 Y3
y1* y2* y3*
w1 w2 w3
MEASUREMENT MODEL FIGURE 2
Is this model identified? YES, but only with equality constraints!

WILEY-WILEY (1970) SOLUTION: ASSUME EQUAL
MEASUREMENT ERROR VARIANCES OVER TIME!!!! IF
THIS IS THE CASE, WE HAVE A JUST-IDENTIFIED
MODEL WITH 5 STRUCTURAL PARAMETERS AND 1
MEASUREMENT ERROR VARIANCE TO ESTIMATE.
Structural parameters: 2 stability effects and 3 error terms for the Yt

Measurement error variance for the w

Finkel (p. 53-54) goes through some of the (tedious) covariance
algebra to solve for the unknowns. Key results:
Cov( y1* , y3* )

2 = * *
Cov( y1 , y2 )
* *
Cov( y , y )
1 = 1 2
Var( y1* ) Var(w)
* * * *
Cov( y , y )Cov( y , y )
Var(w) = Var( y2 )
* 2 3
* *
1 2
Cov( y1 , y3 )
Pretty straightforward, but note that you first need to solve for
Var(w) in order to arrive at the estimate for 2.

EXAMPLE: PARTY IDENTIFICATION,
2000-2002-2004 NES PANEL STUDY
4.68

S = 4.11 4.75
4.30 4.47 5.42
Var(w) = 4.75 ((4.47 * 4.11) / 4.30) = .48

1 = 4.11/ (4.75 .48) = .96
2 = 4.30 / 4.11 = 1.05

Can also calculate the reliability of each indicator of y, following
the formula given on slide 63 above:
Var(Yi ) Var( yi* ) Var(wi ) Var(wi )

yy = *
= *
= 1
Var( yi ) Var( yi ) Var( yi* )
11 = (4.68 .48) / 4.68 = .90

22 = (4.75 .48) / 4.75 = .90
33 = (5.42 .48) / 5.42 = .91

Notes on Example
We have about 90% true score variance in the indicators and 10% error variance.
This shows very high reliability of the partisanship variable, and means that using
OLS on the fallible indicators would not be so bad. Still, correcting for
measurement error is necessary to yield unbiased estimates of the .
Once measurement error is taken into account, there is nearly perfect stability of
partisanship over time. But we have a stability coefficient of 1.05 for waves 2 and 3,
which means we probably have something wrong with the model.
The applicability of the Wiley-Wiley procedure rests on the key assumption of equal
measurement error over time. This is plausible in some cases, but it is sometimes the
case that errors tend to shrink in systems over time as individuals/units learn and
respond more consistently to stimuli in repeated measures. Implausible parameter
estimates may be an indication of violation of this assumption.
A closely related procedure for three-wave single indicator models was developed by
Heise (1969). In that procedure, we use standardized variables for y, so we have 3
known correlations to work with. The model is identified by assuming equal
reliabilities for the y indicators (as opposed to equal error variances in Wiley-Wiley)

1 2
PID2000 PID2002 PID2004

STATA set-up with Wiley-Wiley constraint
on measurement error variances
pid2000 pid2002 pid2004
3 wileywiley 4 wileywiley 5 wileywiley
Results
Notes:
1 .26
2 .27
Zero df, chi-square=0.0
Estimated error variance of single
1
indicator of PID=.48, as we
.98
PID2000 PID2002 PID2004
4.2
calculated manually earlier 1 1 1
Estimated stability=1.045 from pid2000 pid2002 pid2004

wave 2 to wave 3, implausible (and 2.9 3 3
rounded here to 1.0, so be careful 3 .48 4 .48 5 .48
on this!)

Additional Notes on Measurement Error Models
How to include measurement error in more elaborate structural
models? Two choices:
Estimate measurement error separately, and then plug in estimated values as
fixed parameters in subsequent structural runs. In this case, we could take
our estimate of .48 for the error variance, and use this value whenever we
model Party ID in the cross-lagged or synchronous effects models
Estimate unconstrained (free) measurement and structural parameters
simultaneously
Former method is simpler and more stable, less prone to problems with
implausible estimates or lack of convergence; latter method more defensible
statistically because of a more properly specified model
With more waves of observations, can relax Wiley-Wiley
assumptions. In a four wave model, e.g., the inner elements of the
measurement error variances are identified without constraint

6 7
app2000 app2002 app2004
STATA estimation of three-

PID2000 PID2002 PID2004
wave single indicator panel
1 2
model, measurement error in
pid2000 pid2002 pid2004 PID only, fixed at .48
3 .48 4 .48 5 .48
6 .69
7 .38
1.5
app2000 app2002 .44 app2004
1 .1 2.2 .73
.65 .0068
.043
.48 .43 .46
.47
.94 .97
PID2000 PID2002 PID2004
1
.95 .95 .96

1 .058
2 .048
pid2000 pid2002 pid2004

1.4 1.3 1.2
3 .1 4 .099 5 .087
LR test of model: chi2(4) = 10.15, Prob > chi2 = 0.0380

Correcting for Measurement Error in Panel Models:
Multiple Indicators
The last strategy for dealing with measurement
error is to add additional indicators of each latent
variable. If multiple indicators are available, this
strategy is almost always the best, because the
additional variances and covariances will often allow
identification of measurement models without
constraint, and because, logically, having more than
one indicator of a latent variable should provide a
more precise estimation of that variable for use in
subsequent structural modeling.

1
Eval1 Eval2
1 1
Moral2000 InTouch2000 Moral2004 InTouch2004
2 3 4 5
MEASUREMENT MODEL FIGURE 3:

Two Wave, Two Indicator Model

Two indicators of the latent variable at each point in time.
First, we scale the latent variable to equal the scale of one of the
indicators by setting one of the factor loadings at each wave to be
equal to 1. STATA does this automatically by setting the loading of the
first indicator to 1; you can change this if youd like. The other factor
loading will be free to vary. Our unknowns total 9:
1
2 variances of the latent variable disturbances
2 factor loadings (the regression coefficients of Y1 and Y2 on the second
indicator at times 1 and 2, respectively)
4 error variances of the observed indicators
Number of knowns? 4*5/2=10 variances and covariances of the y*. So
we have an overidentified model!
Moral: With multiple indicators you only need two waves to
identify measurement models, as opposed to three waves with
single indicator models.

Fit statistic Value Description
1 .097
Likelihood ratio
p > chi2 0.006
.97 p > chi2 0.000
Eval1 Eval2
.33
Population error
RMSEA 0.135 Root mean squared error of approximation
.77 1 1.1
1 90% CI, lower bound 0.058
upper bound 0.233
pclose 0.037 Probability RMSEA <= 0.05

3.1 2.7 Information criteria
2.9 2.7
AIC 3645.608 Akaike's information criterion
BIC 3695.798 Bayesian information criterion
2 .35 3 .67 4 .52 5 .7
Baseline comparison
Size of residuals
1 .24
unstandardized
.87
Eval1 Eval2
1
.69 .47 .66 .64

standardized
3.2 2.4
3.6 3
2 .52 3 .78 4 .57 5 .58

Notes on Multiple Indicator Models
One conspicuously absent causal effect in this model is a free covariance term
between the measurement errors of a given indicator with its own value over
time. But should not the errors of measurement for indicator 1 at time 1 be
related to the errors of measurement for indicator 1 at time 2? Such a
correlation would be induced, for example, if the indicator was a measure not
only of Y but also of another construct that was unobserved in the model. It
could also be induced through method effects, or interviewer effects that
are consistent across time but are unrelated to the Y
It can be argued that these kinds of correlated measurement errors are essential
to proper estimation of the causal effects in the model, notably the stability
effect 1, which, if these effects are ignored, will be overestimated to the
extent that the correlated errors are not zero and thus account for some of the
over-time covariation between the indicators. (We may also view the model
2 for the model without correlated errors as providing some information
about the plausibility of the restrictions that they are zero, although there are
many other restrictions in the model as given).

However, including these correlated errors renders the model
underidentified, as there are 11 unknowns and only ten knowns.
What can you do?
Add exogenous background variables to the model. As long as these
exogenous variables affect both Y1 and Y2 the model will be identified
without constraint.
Add additional indicators the 2 wave, 3 indicator model is identified without
constraint.
Add additional waves of observations the 3 wave, 2 indicator model is
identified, though not the wave 1-3 error covariances without some constraint.
With more waves of measurement you can also include error covariances
between indicators within a given wave. So measurement errors1 and 2
might be related because they are questions that are right next to one another
in a survey, independent of their respective success in serving as indicators of
Y. More waves equals more covariances and more flexibility in estimating
these effects. There are dangers, though, in overfitting error terms because
they may not replicate across other data sets.

male
1
educ
Eval1 Eval2
1 1
2 3 4 5
.12 .9
male
1 1 .46
2.9
.048
educ -.12
1
Notes: .13
chi2(3) =3.45, P> chi2 =0.3271 .0016 Eval1

.73
Eval2
1
Excellent fit to the data
Much lower estimated stability .69 .47 .58 .74
through inclusion of the

correlated error terms Moral2000
3.6
InTouch2000
3
Moral2004
3
InTouch2004
2.1
2 .52 3 .78 4 .67 5 .46
.13
.22

And of course it is straightforward to estimate measurement error along with
structural effects between variables in the same model (especially so in STATA!!!)
1.4
.11 .22
.91 .084
pid2000 pid2004 6 .25
1 .77 .24
male .47
1 .63 .031
2.9 -.0043
.011 .25
educ -.11 .15
1
1
Strong effect
.06 .35
.016 Eval1
1
.62
Eval2 from PID onto
.6
Eval (.25),
.65 .5 .7
weaker but still
Moral2000 InTouch2000 Moral2004
2.9
InTouch2004
2.1
significant effect
from Eval to
3.6 3
4 5
PID (.15)!
.51 .64
2 .58 3 .75
.13
.11
Two-wave cross-lagged model with multiple indicators in Eval, perfect measurement

assumed in PID.
Chi2(7) = 14.07, Prob > chi2 = 0.05
Not bad!! But remember, the structural portion of the model is saturated (as before),
so nothing really revealed here from the chi-square in that sense
There is interesting discussion in the measurement error literature regarding the
stability of latent variable measurement structures. It is often argued that the
effects from to a given indicator must be the same over time, or else the latent
variable is not the same latent variable over time. That is, 21= 42 (and if there
was a third wave it would also be equal to 63). This is the assumption of
measurement invariance that many argue is essential for proper causal inference.
This constraint can easily be imposed via an equality command. It is recommended
to start with this model and try not to reject it and then no one will criticize you for
it. The philosophical issues are not settled, though, so you could possibly argue your
way to a position whereby is the same but its components change their weights
from time to time.
Estimating the reliabilities of indicators in multiple indicator models is
straightforward: use 1 Var ( i ) as before.
Var ( yi )
Or, in a standardized solution (where y and are both standardized), the reliability
of yi will be 2.

Panel Data Gothenburg - Sessions 3-4.structual Equation Panel Models

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Panel Data Gothenburg - Sessions 3-4.structual Equation Panel Models

Uploaded by

Copyright:

Available Formats

Analysis of Panel Data:

Structural Equation Panel Models

Sessions 3-4: 16 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

How to estimate the structural parameters in this model?

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

So with SEMs we move from a world where we care only about

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

What are the knowns in this system? The observed

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

These are two bivariate regression equations, so in

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Does the model fit the data?

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Observed Correlations Model Implied Correlations

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Basic idea: Find the parameters that generate an implied population

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Note the baseline model has 3 df since there are 3 observed

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

So the CFI is (2288.51-109.77)/2288.51=.952

which equals (762.8-109.76)/761.8=.857

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

SEM Diagram in STATA with

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

(a) Y2 = 1Y1 + 2 X 1 + 1 "static score"

Goal: Compare relative direction and magnitude of cross-lagged effects to

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Systems stabilize at equilibrium points only if 1<1

And since the following is true,

Or, in terms of X and Y :

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

1.4 .79 >

cov(e.pid2002,e.app2002) 77.161 1 0.00 .4519648 .3233477

EPC = expected parameter change

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

pid2000 pid2002 pid2004

app2000 app2002 app2004

Flexibility in model and lag specification increases substantially

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

pid2000 pid2002 pid2004

app2000 app2002 app2004

Wave 1-Wave 3 PID Error pid2000 pid2002 pid2004

Term Covariance Added educ

app2000 app2002 app2004

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

Involve reciprocal causality between variables at a given point in

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

So X1 is related to 2, hence endogenous! Same with Y1 and 1 in

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015

General Moral: we always estimate causal effects within the

Analysis of Panel Data, University of Gothenburg, 15-18 June 2015