Model Identification

7.1. Introduction

▾
7
Model identification plays a preponderant role in structural equa-
tion modelling. Technically speaking, the underpinning rationale
behind model identification is to sieve out unique values for para-
meter estimation. Two tests are widely applied to test for model
identification.

7.2. Types of Parameters in SEM

There are two types of parameters: namely, fixed and free. Fixed
parameters are never estimated from the data since their values
are fixed to be either zero or one. To set up a scale for each latent
variable, there is the need to fix the variance of each latent variable
to one or fixed the value to one of one parameter. Free parameters
are estimated from the data. Figure 7.1 depicts the types of para-
meters used in SEM.

51
 Types of parameters under SEM

Fixed Free

Fix variance of each latent variable to 1 * Parameters that are being estimated,
Fixed the value to one of one parameter namely, the order condition and the
condition

Figure 7.1: Types of Parameters under SEM.

7.3. Types of Identification

Types of model identification

Exact identification Overidentification

Under-identified
Unique value can be obtained No exact solution, a value for
for the parameters one or more parameters can be
Unique values for the obtained in multiple ways
Degrees of freedom = 0
parameters cannot be obtained Degrees of freedom > 0
from the observed data
Degrees of freedom = Negative

Figure 7.2: Types of Model Identification.

A model is identified if it is feasible to derive a unique solution for

every parameter. An identified model must have non-negative
degrees of freedom, that is, the number of estimated parameters
should be less than or equal to the number of data points desired
from the sample covariance matrix, under the famous order
condition. Figure 7.2 demonstrates the distinct types of model
identification.

52 Applied Structural Equation Modelling for Researchers and Practitioners

 If a model is overidentified, it means that the number of equa-
tions exceeds the number of unknowns so that the system of equa-
tions is solvable in more than a single way by virtue of the
existence of many possible solutions. The objective is thereby
geared towards the selection of a model that generates the best fit
to the data. Over-identified models are most susceptible to be of
more substantive interest than just identified models to research-
ers. Under SEM, overidentification can be induced via setting
some parameters to be fixed. Typically, the parameter values are
set to zero. A second way to induce overidentification is to impose
a one-way causality so that the reverse causality value is set to
zero. The ideal solution for researchers is to have an overidentified model
as this offers them the feasibility to eventually select that model which
unleashes the best fit to the data. The degrees of freedom is positive
for an overidentified model.
If a model is just identified, it means that the model is saturated
with degrees of freedom being tantamount to zero. Under just
identified model, there is only one unique solution with the model
always generating a perfect fit to the data. The degrees of freedom
is equal to zero for a just identified model. Just identified models
have zero degrees of freedom, because the number of estimated
paths equals the number of elements in the covariance matrix and
hence generates a perfect fit to the data.
If a model is under-identified, this implies that the number of
unknowns exceeds the number of equations so that there is no
unique solution. Moreover, the model parameters are not uniquely
identified. The degrees of freedom is negative in the case of an
under-identified model.
Covariances can prevail not only among observed variables but also
among latent variables. In a parallel manner, there are variances not
only for the observed variables, when they capture the latent variables,
but also for the latent variables which constitute dependent variables.

M ode l Ide ntif icat ion 53

7.4. Examples of Model Identification Explained
7.4.1. Under-Identified Model

d Independent
Independent A B
variable
variable

a
C
b

Latent dependent
variable

Figure 7.3: Under-Identified Model.

In practice → need an overidentified model.

The channel to check for identification is the degrees of freedom.

Number of observed variables = 2
Hence, the number of unique elements in the covariance matrix
is 2ð2þ1Þ
2 ¼3

Number of model parameters

= Number of regression coefficients
+ Number of variances + Number of covariances
Number of regression coefficients
= 2 (based on single-headed arrows)
Number of variances
= Variances of independent observed variables
+ variances of error terms
=2+1=3
Number of covariances among independent variables = 1
Thus, the degrees of freedom
= number of unique elements in the covariance matrix
minus the number of model parameters
= 3  6 = 3 < 0 (Underidentified model; df < 0)

54 Applied Structural Equation Modelling for Researchers and Practitioners

It is plain obvious from the above exercise that there are two ver-
sions of the error terms. First, an automatic error term follows
when gauging the effects of the independent variables on a given
dependent variable. Second, error terms emanate from observed
variables which capture latent variables. Figure 7.3 depicts an
example of an under-identified model in SEM.

7.4.2. Overidentified Model (The One to Be Used)

Indicator
Indicator d B variable
A
variable

b
a Mediating latent
C variable
e

D E Independent
variable
Independent
variable

Figure 7.4: Over-Identified SEM.

4 manifest known variables: A, B, D and E. Figure 7.4 shows

example of an over-identified SEM.
Number of variables: 4, so number of unique elements in the
covariance matrix is 4ð4þ1Þ
2 = 10
Number of model parameters
¼ Number of regression coefficients
þ Number of variances þ Number of covariances
Number of regression coefficients
¼ 4 ðbased on single-headed arrowsÞ

M ode l Ide ntif icat ion 55

 Number of variances ¼ Variances of independent observed
variables þ variances of error terms
¼2þ1¼3
Number of covariances among independent variables = 1 (double-
arrowed among indicator variables)

Thus; the degrees of freedom ¼ number ofunique elements

in the covariancematrix minusthe number of model parameters
¼ 10  8 ¼ 2 > 0 ðover-identified model; df > 0Þ

7.5. Model Identification: Both Measurement

and Structural Equations
Focus is being laid on an SEM which has the measurement and
the structural equations formulated with an unknown parameter
vector θ. The definition of identification is based on Σ(θ), the popu-
lation covariance matrix of the observed variables in y. A model is
said to be identified if for any θ1 and θ2,

∑(θ ) = ∑(θ )

Implying that θ1 = θ2

However, SEM comprises of both the measurement and the structural

models so that SEM is identified if both its measurement equation and
structural equation are identified.
m(θ): measurement model
s(θ): structural model
Measurement model is identified if for any θ1 and θ2, m(θ1) = m
(θ2) so that θ1 = θ2
Structural model is identified, if for any θ1 and θ2 , s(θ1 ) = s(θ2 ),
implying that θ1 = θ2

56 Applied Structural Equation Modelling for Researchers and Practitioners