The integrated trait–state model

Abstract

It has been acknowledged that both trait and state contribute to psychological measurements. However, existing structural equation models for disentangling these sources of variability are based on assumptions that are not tenable in the light of empirical results. A new model is presented, termed the integrated trait–state (ITS) model, which both decomposes state and trait variance and allows one to test the assumptions that underly existing approaches. This is illustrated with an empirical example. The relationship between the ITS model and other analytic approaches as well as conceptual models of traits and states are discussed.

Introduction

Traditionally, traits have been an important object of psychological research, especially in the realm of personality and ability. The trait concept and its utility have also received sharp and lingering criticism at various times even from those committed to the study of personality (e.g., Mischel, 1968, Pervin, 1994). One of the major concerns has arisen in longitudinal research where it was shown that traits are not as stable as they were once thought to be both across occasions and over situations. Also, the prediction of future behaviors based on current traits turned out to be less than impressive. Consequently, social psychologists, for instance, argued that not person (i.e., trait-like) characteristics but situational characteristics should be a key focus of psychological research. Still other writers claimed that both person and situation characteristics and their interactions are fundamental to understanding cross-situational stability and variability in a person’s behavior (Buss, 1989, Epstein, 1984, Smith, 1988). This interactional view is closely related to the trait–state distinction, in which a trait is considered to be a person characteristic that remains stable across time and situations, and a state is considered to reflect a person’s adaptation to a particular situation. From this it follows that when individuals are measured at a given occasion it is likely that variation in both trait and state contribute to the variation in observed behavior.

By acknowledging that traits and states may be confounded in psychological measurements, how to disentangle them becomes a salient issue. Several structural equation modeling (SEM) approaches to this problem have been suggested. For instance, Kenny and Zautra (2001) have introduced the stable trait, autoregressive trait and state (STARTS) model. While this model can be used to model the relatedness of observations over time, it cannot be used to investigate similarities and differences across people’s state structures, or across the trait and the state level since it is an univariate technique. A multivariate longitudinal SEM approach is the latent state–trait (LST) model developed by Steyer and his colleagues (Schmitt and Steyer, 1993, Steyer et al., 1990; for extensions see Cole et al., 2005, Schermelleh-Engel et al., 2004, Tisak and Tisak, 2000), which allows for the simultaneous modeling of trait-like and state-like variability in a factor model. However, it is based on the assumption that the factor structure underlying these two kinds of variability are identical. This implies that the factor structure of states must be identical across individuals. Although there exists some evidence for this assumption (e.g., Borkenau and Ostendorf, 1998, Garfein and Smyer, 1991), there are many studies that led to the conclusion that individuals’ state structures are characterized by important idiosyncracies (Hamaker et al., 2005, Hooker et al., 1987, Hurlburt and Melancom, 1987, Quinn and Martin, 1999, Shifren et al., 1997, Zevon and Tellegen, 1982).

In the current paper, a new SEM-based technique is presented to overcome the limitations of these previous techniques through combining several existing ideas from the trait–state literature. We illustrate the use of this new approach with an empirical example. At the end of this paper we discuss ways in which this new approach can be extended, and how it is connected to other conceptualizations of traits in the personality literature. For clarity we begin by indicating how the terms trait and state are used throughout this paper.

The distinction between traits and states dates back at least to the beginning of the Christian era (Eysenck, 1983). It has been cast in terms of stability versus change (Hertzog & Nesselroade, 1987), consistency versus change (Cervone, 2004), consistency versus discriminativeness (Funder, 1994), invariance versus variability (Mischel, 2004), disposition versus dynamics (Mischel, 1973), and person versus situation (Meijer, 1994, Steyer et al., 1999). Although few (if any) will hold the extreme viewpoint that trait level never changes across the individual’s life-span, most will probably agree that stability in some form (see e.g., Mortimer, Finch, & Kumka, 1982) is the most distinctive feature of a trait. Traits have been defined as relatively stable, interindividual differences in proneness (Eysenck, 1983, Spielberger and Sydeman, 1994), tendency (Spielberger & Sydeman, 1994), style (Mischel, 1968), or disposition (Block, 1993, Forgays et al., 1997) to behave (Block, 1993, Mischel, 1968), feel (Spielberger & Sydeman, 1994), or think in certain ways (Goldberg, 1994, Pervin, 1994, Spielberger and Sydeman, 1994). A straight forward application of this idea is to consider an individual’s mean over time (which is cleared from situational influences) as his/her trait score (Buss, 1989, Epstein, 1980). While this is the way in which trait is conceptualized in the following sections, we discuss a more general conceptualization in the final section of this paper.

Defining traits as invariant across time and situations implies that traits are time-invariant interindividual or between-person differences. In addition, we can distinguish between two kinds of time-varying interindividual differences, both of which are depicted in Fig. 1. The first is intraindividual or within-person change, also referred to as trait change (Cattell, 1978), which has been defined as relatively slow changes that are (more or less) irreversible, reflecting processes such as maturation, learning, and progressive organic damage (Nesselroade, 1991). The second is intraindividual variability, also referred to as state variation (Nesselroade, 1991, Nesselroade, 2001), and it is defined as relatively rapid and reversible variability that takes place around the intraindividual’s mean or trend. This implies that the intraindividual mean represents a person’s trait-level, while a changing mean represents trait change. The discrepancy between the mean or trend and the actual observation is the state.¹ These states may be associated with the exogenous environment, such as the social and physical situation, or the endogenous environment, such as physiological, emotional, and cognitive processes taking place within the individual. Hence, what is often referred to as measurement error and discarded, we refer to as state and consider potentially meaningful.

Clearly, the concepts of intraindividual stability, variability, and change are closely related. In this paper, we focus on intraindividual stability (i.e., trait) and variability (i.e., state) because we hold the opinion that understanding the trait–state issue is a prerequisite for studying the dynamics of trait change.

Factor analysis and principal component analysis are two techniques that are widely used in trait and state research to determine a reduced number of dimensions underlying the observed set of variables. Both techniques are based on analyzing the covariance structure of the observed variables. Thus, to explain how it is possible that dimensions found at the interindividual level are not necessarily replicable at the intraindividual level, we need to look at the covariance between variables at these different levels. To simplify matters, we look at a bivariate example.

Suppose we take shyness and sociability: Shyness refers to inhibition and feelings of insecurity when with others, while sociability is a preference for being with others (Buss, 1989). Shyness and sociability are negatively correlated in the population implying that scoring high on shyness is associated with scoring low on sociability and vice versa, as illustrated in Fig. 2a. However, this is a relationship at the population level, which does not have to hold for each individual case: An individual may deviate from this population pattern by scoring low on both shyness and sociability (see Fig. 2a).

But there is another way in which the individual may differ from the pattern found in the population: The negative relationship between shyness and sociability may prove nonexistent or even reversed at the intraindividual level. To determine the intraindividual relationship between shyness and sociability we have to measure these same two variables repeatedly in the same individual. Based on these measurements we can determine the intraindividual correlation, which provides insight into the way different states covary within a particular person. In Fig. 2b we illustrated the relationship between shyness and sociability for a particular individual who is quite high on shyness while low on sociability compared to others. Within this individual a positive relationship between shyness and sociability emerges. A possible explanation for this is that when this person feels an increased desire to participate in social interaction, this also raises his/her social anxiety levels. Other individuals may be characterized by other intraindividual relationships, for instance, no relationship or a relationship which is more in agreement with the relationship found at the interindividual level.

We want to stress here that we do not expect that for many psychological phenomena the intraindividual relationship between two variables is opposite to the relationship at the interindividual level: The above is simply meant to exemplify that relationships between variables may differ at different levels. This fact has been recognized by many (e.g., Baldwin, 1946, Borsboom et al., 2003, Epstein, 1994, Grice, 2004, Hamaker et al., 2005, Lamiell, 1990, Nesselroade and Molenaar, 1999, Schmitz and Skinner, 1993, Von Eye and Bergman, 2003). When the relationships between variables are identical at both levels, this has been referred to as local homogeneity, while a difference between the structures at different levels has been coined local heterogeneity (Borsboom et al., 2003). This notion is crucial for the model developed in this paper, which allows individuals to differ from one another with respect to their structure of intraindividual variation: As a result, this model offers the opportunity to test the assumptions underlying other approaches, i.e., whether there is a universal state structure that applies to all, and whether or not it coincides with the trait structure. As such it allows us to investigate whether constructs are locally homogeneous or heterogeneous (Borsboom et al., 2003).