You are on page 1of 3

QMM: Exercise Sheet 8 - Structural Equation Model:

Structural Regression

Fabien Baeriswyl, Jérôme Reboulleau, Tom Ruszkiewicz

Exercise 1. Consider the following graph:

In the above, we use Anomia and Powerlessnes as subscales of an Alienation scale. Social Anomia refers to a
weakening of cultural values and social norms, while Social Powerlessness refers to the sensed ability to control
outcomes. In turn, Social Alienation is ”a condition in social relationships reflected by a low degree of integration
or common values and a high degree of distance or isolation between individuals”.
Note also that in the graph above, Education is measured in years; SEI represents a soccioeconomic index based on
the respondent’s occupation while SES stands for socioeconomic status. Note that there two years of observations:
1967 and 1971 and related variables are emphasised by ”67” or ”71”, respectively.
a) In the graph, identify the latent and observed variables.
b) Determine the model degrees of freedom.
c) Determine your model equations and create the model in R.
d) Often, it is impossible to simply access the data. Rather, we have the Covariance matrix available to us. You
can find how to construct it with the help of the lavaan package under the CovMat.R file. Fit the model in R
using this Covariance matrix with the sample.cov parameter, specifying 932 sample observations.

1
e) Analyse your results and comment about the model quality.
f) Define the Null Model and R and manually check the values obtained for the CFI and TLI from the previous
point.
g) Respecify your model by eliminating all covariance terms. How could this new model be improved? (Hint: use
the modificationIndices() function in R.)
h) Check the estimated improvement on the χ2 statistic by adding one path suggested in 1g). Comment on the
quality of the improvement.
Exercise 2. Calsyn and Kennly (1977) recorded the values of the following six variables for 556 eighth-grade
students:
ˆ SCA: self-concept of ability;
ˆ PPE: perceived parental evaluation;
ˆ PTE: perceived teacher evaluation;
ˆ PFE: perceived friend evaluation;
ˆ EA: educational aspiration;
ˆ CP: college plans.
Calsyn and Kenny (1977) postulated that two underlying latent variables, ability and aspiration, generated the
relationships between the observed variables.
The first four variables were assumed to be indicators of ability and the last two indicators of aspiration. The latent
variables, ability and aspiration, are assumed to be correlated.
a) Draw the DAG corresponding to the description above.
b) State the model’s equations.
c) Compute the model degrees of freedom in R.
d) The covariance matrix related to the above is given in the CovMat2.R file. Use it to fit the model.
e) Refit the model but fix the latent variable variances to 1.
f) What is the effect of putting a measurement error in the latent variables?
Exercise 3. This simple exercise, based on simulated data, will show you a nice (and quick) justification of SEM.
We want to generate some bivariate data (y1 , y2 ) that are, by construction, correlated. From these variables, we will
derive x1 , x2 , x3 , x4 that will be considered observed variables. We will try to determine the correlation between y1
and y2 that we know by construction but that we will assume unknown.
a) Using the MASS package and more specifically the mvrnorm() function, generate 100 (y1 , x2 ) bivariate data
from a bivariate normal disribution with corrrelation 0.6. Then, generate x1 and x3 by adding to y1 and
y2 , respectively, an exponential noise of rate 1 (use the rexp(n, rate) function in R to generate n random
exponential observations with rate=rate) and generate x2 and x4 by adding to y1 and y2 , respectively, an
exponential noise of rate 2. Create a dataset with the variables x1 , x2 , x3 , x4 .
b) Fit a SEM in trying to model the latent variable y1 by the observed variables x1 and x2 and the latent variable
y2 by the observed variables x3 and x4 . What can you say about your fit? About the estimated covariance
coefficient?

2
c) Should you add covariances between, on one side, x1 and x2 and on the other side, x3 and x4 ? Justify.
d) Add a constraint specifying that b2 < exp(b4 ), where b2 is the coefficient corresponding to x2 and b4 is the
coefficient corresponding to x4 . What do you observe?

You might also like