Deconstructing the construct: A network perspective on psychological phenomena - Schmittmann - 2013 - Article

Introduction

In current psychology, theorizing and research is dominated by two conceptualizations of the relationship between psychological attributes (such as depression) and observable variables (feeling sad, losing weight). The first conceptualization is defined as the reflective model. The reflective model states that the attribute is the cause of the observed variables (so, depression is the cause for the symptoms). The second conceptualization is the formative model. In the formative model, the observed scores define the attribute. An example of a formative model is socio-economicstatus (SES), which is viewed as the combined effect of variables like education, job, salary and neighborhood. The authors of this article argue that these two conceptualizations are not sufficient for describing all the relationships between psychological attributes and observable variables. They advocate for a different kind of conceptualization, in which psychological attributes are conceptualized as networks of directly related observable variables. 

Reflective and formative models

Problems with the reflective and formative conceptualizations

Reflective and formative measurement models have often been the source of discussions. These discussions are often focused on the question whether there are reasons to favor one model over the other model. For example, some researchers questioned whether formative models should even be used, and other researchers stated that reflective models are used when there should be formative models used. 

The authors of this article say that the causal relationships that formative and reflective models try to depict, are problematic in psychological testing. They discuss three particular problems: the role of time, the inability to articulate causal relations between construct and observables in terms of processes, and the subordinate treatment of relations between observables.

The role of time

When we talk about causality, then this means that causes precede the effects. However, in psychometric models such as reflective and formative models, this time is not explicitly represented. Therefore, it is unclear whether the latent variables relate to the observables. 

Inability to articulate processes

Establishing causal relationships is very important for science. Normally, after discovering a causal relationship, one goes on to explain the process of the precise mechanisms that lead to this relationship. So, for example, when researchers discovered that there is causal relationship between smoking and lung cancer, they studied what elements of the causal factor (tobacco smoke) result in the effect (lung cancer). The authors point that this kind of explanation is lacking in psychological research. So, in psychological research, it is concluded that neuroticism causes worry about things, but it is never explained how neuroticism causes this worry. However, this is understandable, because it is very hard to understand how these effects arise. The authors state that this is due to the fact that constructs in psychology (neuroticism) are not empirically identifiable. For example, 'general intelligence' is not found somewhere in the brain. 

Relations between observables

Reflective models have the assumption that there is no direct causal relation between observables. In the formative models, relationships between observables that are not accounted for by the latent variables are treated as nuisance. However, the authors state, it is very likely that there are causal relations between observables in many psychological constructs. These causal relations may even be the reason for some constructs, such as SES, which is summarized as a group of variables. 

The network perspective: constructs as dynamical systems

The authors state that the variables that are seen as indicators of latent variables, should not be viewed merely as indicators. Instead, they should be viewed as autonomous causal entities in a network of dynamical systems. So, instead of assuming a latent variable, one should assume a network of directly related causal entities. When assuming the latter, the three problems described are avoided.

For example, let's look at the criteria for major depressive disorder. The symptoms of major depressive disorder include: "lack of sleep", "fatigue", and "concentration problems". In empirical research, these criteria are scored and then added to form a total score, which then reflects "depression". However, this practice ignores that there are direct relations between these symptoms (so, lack of sleep leads to fatigue, which leads to concentration problems). In a dynamic scheme, these elements are connected causally in a network of autonomous causal entities. When one has constructed this kind of network, he or she can use techniques to visualize the system. So, in a network perspective, a construct is seen as a network of variables. A change in one variable causes a change in another variable. So, when one wants to study a psychological construct, he or she needs to study the network of this construct. This study would focus on a) the network structure and b) network dynamics. The relationship between variables should not be seen as a measure of the construct, but as part of the construct.

Dynamical systems

The dynamical systems is a way to study the behavior of a network of interconnected variables over time. In psychology, this has often been applied to cognitive processes, to the construct of intelligence and to the area of developmental psychology. A dynamical system is said to change its state, based on equations that describe how the previous state determines the present state (so, how the variables influence each other). Attractor states are important in dynamical systems. For example, a depression network may have two attractor states: a disordered and depressed state, and a healthy state. Things such as stressful life events may lead a person to shift from the healthy state to the disordered and depressed state. 

Causal inference

A method to construct causal systems is the use of inference: a statistical method. These methods work through the detection of conditional independence relations. 

Network analysis

When the network structure has been determined, the networks can be analyzed further. There are many network structure analysis methods (for instance in R). With the use of such methods, one can examine whether a network has small world properties (high clustering of items with relatively few separating nodes). One can also analyze the properties of individual nodes such as their centrality (how strongly the node is connected to all the other nodes in a network).