LATENT VARIABLE

JUNE 7, 2018
PRASHANT S OJHA
UEMF17017
CONSTRUCTS AND MEASURES
Constructs are that which cannot be directly measured (but we assume exists), where measures are
directly measurable attributes that we assume relate to the construct.
The underlying phenomenon or construct that a scale is intended to reflect is often called the Latent
Variable. In statistics, latent variables (as opposed to observable variables), are variables that are
not directly observed but are rather inferred (through a mathematical model) from other variables
that are observed (directly measured). Mathematical models that aim to explain observed variables
in terms of latent variables are called latent variable models. Latent variables are features that are
not directly provided during the training process. They are not immediately evident in the training
data and are hidden variables. For example, if you are designing a face recognition model, the
training data has pixels and the label of every image. It doesn't have any information about scale
and pose that can be observed in the training data directly. These variables that are useful in
predicting the class are hidden or latent variables. For e.g., if we ask a group of shoppers’
perceptions rather than aspects of the store itself (which might be more easily assessed by direct
observation). How important the distinction is between assessing the perceptions of a respondent
with regard to some external stimulus (e.g., perceptions of the store), as opposed to characteristics
of the external stimulus (e.g., the store itself), will depend on the specific circumstances and goals
of the assessment.
A causal relationship between a latent variable and a measure implies certain empirical
relationships.

PATH DIAGRAM AND PATH ANALYSIS

Path diagrams are a method for depicting causal relationship among variables. Path analysis is a
data analytic method and doesn’t have more general utility as a means of specifying how a set of
variables are interrelated. These diagrams adhere to certain conventions. A straight arrow for e.g.,
drawn form one variable label to another indicates that the two are causally related and that the
direction of causality is as indicated by the arrow. Similarly, the absence of an arrow also has an
explicit meaning- namely, that the two variables are unrelated. Path diagrams also is a method of
representing error, which is usually depicted as an additional causal variable. How scale items are
causally related to a latent variable can be shown by path diagrams. A path diagram with path
coefficients can be used to compute correlations between variables. The numbers along the paths
are standardized path coefficients. Each one expresses the strength of the causal relationship
between the variables joined by the arrow. the relationship between path coefficients and
correlations provides a basis for estimating paths between a latent variable and the items that it
influences.
MEASUREMENT MODEL
Measurement models refer to the implicit or explicit models that relate the latent variable to its
indicators. Several basic questions require consideration prior to building a measurement model:
A. Are there causal or effect indicators?
B. Are there multiple or single indicators of the latent variable?
C. Is the latent variable continuous or noncontinuous?
D. Are the indicators continuous or noncontinuous?
E. Does more than one latent variable influence the indicator?
The first question concerns whether the indicators influence the latent variable (causal indicators)
or whether the latent variable drives the indicators (effect indicators). The second characteristic
of measurement models is whether there are multiple indicators of a latent variable or just one.
Multiple indicators are generally preferable to single measures. The third and fourth distinctions
from above are whether the latent variables and indicators are continuous or not. Continuous refers
to whether the latent or indicator variable is best thought of as approximating a variable that has
infinite gradations of magnitude or not.

PARALLEL TESTS
Parallel testing means testing multiple applications or subcomponents of one application
concurrently to reduce the test time. Parallel tests consist of two or more parts (projects or project
suites) that check different parts or functional characteristics of an application. These parts are
executed on individual computers simultaneously. The ability to test more than one application
part "in parallel" allows you to significantly reduce the test time and test your applications more
efficiently. The parallel model. The parallel model is the most restrictive measurement model for
use in defining the composite true score. In addition to requiring that all test items measure a single
latent variable (unidimensionality), the parallel model assumes that all test items are exactly
equivalent to one another. All items must measure the same latent variable, on the same scale, with
the same degree of precision, and with the same amount of error. The rationale underlying the
model of parallel tests is that each item of a scale is precisely as good a measure of the latent
variable as any other of the scale items. The individual items are thus strictly parallel, which means
that each item’s relationship to the latent variable is presumed identical to every other item’s
relationship to that variable and the amount of error present in each item is also presumed to be
identical. Do not confuse parallel tests with distributed tests. Both of these kinds of tests consist
of several parts that are performed on separate computers. However, distributed testing also
presupposes that test parts interact with each other during the test run, while the parallel testing
approach lacks interaction between test parts.

TAU-EQUIVALENT MODEL
The tau-equivalent model is identical to the more restrictive parallel model, save that individual
item error variances are freed to differ from one another. This implies that individual items measure
the same latent variable on the same scale with the same degree of precision, but with possibly
different amounts of error (Raykov, 1997a, 1997b). All variance unique to a specific item is
therefore assumed to be the result of error. The tau-equivalent model implies that although all item
true scores are equal, each item has unique error terms:
Xik = Ti + Eik.
The essentially tau-equivalent model. The essentially tau-equivalent model is, as its name implies,
essentially the same as the tau-equivalent model. Essential tauequivalence assumes that each item
measures the same latent variable, on the same scale, but with possibly different degrees of
precision.

THE CONGENERIC MODEL

The congeneric model is the least restrictive, most general model of use for reliability estimation.
The congeneric model assumes that each individual item measures the same latent variable, with
possibly different scales, with possibly different degrees of precision, and with possibly different
amounts of error (Raykov, 1997a). Whereas the essentially tau-equivalent model allows item true
scores to differ by only an additive constant, the congeneric model assumes a linear relationship
between item true scores, allowing for both an additive and a multiplicative constant between each
pair of item true scores, as shown below:
Xik = [ k + k( i )] + Eik
To identify the model that is with the congeneric model, the path from the latent true variable to
one of the measured items is set to 1, whereas the other paths Graham / Congeneric Reliability 935
from the true variable to the items are left free to be estimated. This indicates that the true scores
of the other items are expressed in terms of the true score of the fixed item. Any of the measured
items can be chosen as the “scaling” variable, with no effect on the outcome of the model.

GENERAL FACTOR MODEL

This model does not assume that only one latent variable is the source of all covariation among
the items. Instead, that model multiple latent variables to serve as causes of variation in a set of
items. The chief benefit is its improved correspondence to real-world data. Structural equation
modeling approaches often incorporate factor analyses into their measurement models ; situations
in which multiple latent variables underlie a set of indicators exemplify the general factor model.
IRT MODEL (ITEM RESPONSE THEORY)
IRT explicitly models firms' and individuals' observable actions in order to measure unobserved,
latent characteristics. IRT models have helped researchers improve measures in numerous
disciplines.
Each of the previously discussed measurement models is part of a nested hierarchy. The congeneric
model shown earlier is the most general, least restrictive model for use in reliability estimation. If
all multiplicative constants in the congeneric model are set to 1 (inferring that item true scores are
measured on the same scale, or have the same standard deviation), we arrive at the essentially tau-
equivalent, or coefficient alpha, model. If all additive constants are then set to 0 (inferring that not
only do item true scores have the same variance, but they are measured with the same degree of
precision, or have the same mean), we arrive at the tau-equivalent model. Finally, if all error
variances are set to equal one another, we arrive at the parallel model, where all observed and
latent variables are equivalent across items.