Department of Psychology, University of Amsterdam, The Netherlands
Complex Trait Genetics, Department of Functional Genomics and Department Clinical Genetics, Center for Neurogenomics and Cognitive Research (CNCR), FALW-VUA, Neuroscience Campus Amsterdam, VU University Medical Center (VUmc), The Netherlands
People differ widely in how they navigate through the landscape of life: some people feel comfortable around others and like to go to parties, whereas others do not. Some worry much and often have trouble sleeping, whereas others rarely experience such problems. Two main challenges in personality psychology are (i) to provide a plausible account of how the coherent ‘organization’ of such behaviours arises within an individual and (ii) to describe and explain the structure of ‘individual differences’ in personality (Caprara & Cervone, 2000). Modern personality psychology has mainly focused on the latter task. Starting with pioneering work of, among others, Thurstone (1934), the dominant doctrines in current personality theory have come to define individual differences in the structure of personality in terms of a number of unobserved trait ‘dimensions’ (e.g. neuroticism and extraversion; Berrios, 1993; Digman, 1990; Goldberg, 1993).
In most interpretations of this concept, consistent differences between people in the behaviour they display are thought to result from underlying differences in these personality dimensions. The interpretation of the term ‘underlying’ is typically borrowed from measurement models in psychometrics, which invoke latent variables—variables that are indirectly measured through a number of noisy indicator variables (i.e. personality inventory items; Borsboom, 2008a). In line with this mode of thinking, the items of a personality inventory are usually considered to be ‘trait measurements’, for example, the item ‘I like to go to parties’ is considered a measurement of the dimension/trait extraversion. Analogous to temperature, which causes mercury to rise and fall in a mercury thermometer, personality dimensions are presumed to cause responses to personality questionnaire items. That is, higher levels of extraversion cause people to make friends more easily and to feel good in the company of others, and these properties are queried in typical questionnaire items. Thus, for example, Alice is not only more extraverted than Bob in the sense that her responses can be ‘described’ by a higher position on an abstract personality dimension (i.e. extraversion); her higher level of extraversion is what ‘causes’ her to like parties better than Bob does. In this way, personality dimensions are interpreted as causes of human behaviour. Perhaps the most outright commitment to this point of view is expressed in McCrae and Costa (2008, pp. 288) who claim that ‘E[xtraversion] causes party-going’.
In the present paper, we challenge this approach to personality. In addition, we offer an alternative perspective. We propose that personality is a system of inter-connected affective, cognitive and behavioural ‘components’. More specifically, we propose that every feeling, thought or act is a potential component of personality if it is associated with a unique ‘causal system’: the pattern of causes and effects that the component exhibits in relation to other components. The component thus must be unique in the sense that its causal system differs from that of other (potential) components. This means that a personality component is, to a certain degree, causally autonomous and, as such, not ‘exchangeable’ with other components. Thus, liking parties is a personality component because it has unique causes and effects on other components (e.g. being interested in meeting new people and not feeling insecure about making a good first impression) that differ from the causes and effects of other components (e.g. starting conversations easily, also an extraversion item, does not necessarily imply that one is interested in meeting new people). To the contrary, making to-do lists that are followed point by point and sorting clothes by colour may not be separate components at the level of personality (i.e. their causes and effects on other components will likely be similar) but two ways of assessing one component, namely liking order. Barring such exceptions, personality components are typically assessed through single items in personality inventories.1 This is because two items that assess precisely the same component will be very highly correlated, which tends to cause problems in typical psychometric analyses (i.e. this will show up as correlated errors).
We hypothesize that such components cannot change independently of one another and, therefore, form a network of mutual dependencies that may alternatively have causal, homeostatic or logical sources. Directional dependencies (typically associated with ‘causality’) will form if one component influences the other but not the other way around: for example, if one cannot plan ahead, it is difficult to meet obligations at work. Bidirectional dependencies will form if two components influence one another (and, as such, create a feedback loop): for example, after a sleepless night worrying, one may feel stressed out and tired the next day; as a result of which, one may not sleep the following night either because of worries about yet another sleepless night. An important special case of feedback involves negative feedback loops that serve to maintain ‘homeostasis’: for instance, after a few sleepless nights, one will ordinarily get so tired that one will start sleeping again (incidentally, if this does not happen, problems are likely to spread to other components, e.g. not being able to concentrate and foul mood; Cramer et al., 2010). Finally, semantically ‘logical’ dependencies will form if two components assess the same or a narrower/broader version of a personality characteristic (which may ultimately but not necessarily result in these two components being merged into one component): for example, liking a clean house and liking a clean desk. We postulate that the resulting pattern of connectivity among such components provides a fruitful avenue into personality research. Also, the dependencies between these components result in a typical network architecture (e.g. being interested in other people and spending time with them are mutually dependent while planning ahead and liking people are not) that can serve as a sufficient explanation of the correlational structures typically observed in personality research (e.g. trouble falling asleep and feeling jittery are more strongly correlated than feeling jittery and liking people).
This opens the perspective of a personality theory that is holistic (i.e. that is about the ‘organization’ of behaviour: network architecture) and that addresses personality at the level of the individual but is nevertheless systematically formalizable through network models. Importantly, this view does not regard personality dimensions as causes of behaviour. We will instead argue that personality dimensions emerge out of the connectivity structure that exists between its components, such that certain components cluster together more than others, with the known personality dimensions as a result.
The structure of this paper is as follows. First, we examine affective, cognitive and behavioural components of personality and argue that a network perspective naturally accommodates mutual (in)dependencies among them (see also Cramer et al., 2010; Schmittmann et al., ; cf. Read et al., 2010, for similar perspectives). We describe the consequences of the network perspective for prominent topics in personality psychology in the subsequent sections of the paper, which deal with the state–trait distinction, the relation of personality to psychopathology and the genetic basis of personality. Throughout the paper, we relate the network perspective to currently dominant trait theories.
HUMANS AS DYNAMICAL SYSTEMS
Few psychologists would challenge the conclusion that human beings are complexly organized. Even the simplest behavioural act (e.g. starting a conversation with a stranger while waiting for the bus) reinforces cognitive schemas (e.g. it provides evidence for the hypothesis that one is capable of starting such a conversation) and affective conditions (if the small talk is successful, this most likely generates a feeling of satisfaction). Because these cognitive and affective components are associated with a class of behaviours in a given situation, they almost certainly serve to sustain the ability and willingness to execute these behaviours when a similar situation is encountered in the future (Mischel & Shoda, 1995). That is, one who has successfully engaged in small talk and enjoyed it is likely to engage in small talk again.
Thus, even this extremely simple example suggests the presence of feedback loops among the components of the personality system, in which behaviour is not just an outcome variable in need of explanation but itself may serve as input to the system (i.e. the behaviour was successful so probably will be executed again under similar circumstances in the future). There most likely are many such feedback mechanisms operating at different time scales, giving rise to a dauntingly complex picture. Thus, Skinner (1987) was definitely on target when he said that human behaviour is ‘possibly the most difficult subject ever submitted to scientific analysis’. In fact, in view of the stunning complexity of the system, it should be considered remarkable that stable behavioural patterns exist at all.
But they do. For some reason, human systems tend to settle in relatively fixed areas of the enormous behavioural space at their disposal, where they are in relative ‘equilibrium’ with themselves and their environments. By equilibrium, we mean a stable state (e.g. Joan is interested in other people and sympathizes with their feelings; as a result of which, she has a job as a social worker) that is not left upon a small disturbance (e.g. one of Joan's clients steals some money from her; after which, she is naturally disappointed in the culprit, but she is still interested in people and their feelings, and she continues her job as a social worker). This definition of equilibrium is analogous to ‘attractors’ in the complex systems literature (e.g. Teschl, 2008).
The idea that human beings strive to survive and reproduce by actively interacting with their environments is an old one and can be traced back to Darwin (1871). In psychology, several scholars have argued for a theory in which human beings are open systems that are constantly searching for equilibrium or a state of homeostasis. Such equilibriums have been argued to exist with respect to components internal to the human system (e.g. in Freudian psychology, id and superego) and with respect to the relation between the human system and its environment (Allport, 1960; Stagner, 1951; Tryon, 1935).
Such states of homeostasis, which we designate to be ‘behavioural equilibriums’, can be achieved and maintained in several ways. For instance, people can and will (consciously or not) seek out environments that match their behavioural repertoire (e.g. Heady & Wearing, 1989; Kendler & Baker, 2007; Kendler, Gardner & Prescott, 2003). For instance, Alice, who loves to go to parties, actively seeks environments that provide many opportunities to party and to meet people who can invite her to parties. Thus, organism–environment feedback loops are important sources of stability because they can serve to sustain behavioural patterns. As a consequence of such feedback-driven selection of environments, however, other behavioural states can also become more difficult to access. This is because they would require different types of environments. For example, Alice cannot both love parties and dislike being around people at the same time. For the former preference to thrive, a socially busy environment with many parties is required, whereas the latter preference would require a more tranquil environment featuring only a limited number of people. Thus, active selection of environments has two important consequences. First, it allows people to settle in a ‘typical’ pattern of behaviour (a behavioural equilibrium, analogous to an ‘attractor’ in complex systems theory) through organism–environment feedback loops. Second, it creates negative dependencies between behaviours that require different environments because people's behavioural options are not inexhaustible: every behavioural act comes at the expense of not performing another and, as such, closes the futures that could have been if another act had been chosen.
It is further characteristic of the behavioural patterns typically studied under the rubric of personality that they can be shaped and maintained in a variety of ways. Thus, people can respond in their own idiosyncratic ways to the situations in which they find themselves. For example, Jane does not like being around people she does not know very well, so when an acquaintance throws a party, she will attend but she will not mingle much and go home as early as is politely possible. However, at a family reunion, she enjoys the company of her close relatives and stays late to catch up with them. These idiosyncratic patterns of situationally dependent responses have been addressed in the cognitive–affective personality system, which we consider to be naturally compatible with a network perspective (CAPS; Mischel & Shoda, 1995, 1998). According to the CAPS model, personality depends not only on the person but also on the environment, that is, one's idiosyncratic way of behaving is stable within environments but variable across environments.
As a result, each person defines a somewhat idiosyncratic equilibrium with his or her environment that is likely to be organized around some properties that play key roles in the individual's cognitive and affective economy, that is, that are important to the person (Cervone, 2005). Because of the connectivity structure of the human–environment system, these properties cannot vary entirely in isolation: one is unlikely to enjoy parties if one does not like people, one is less likely to enjoy company if one is nervous around others, and one cannot be nervous around others if company makes one feel comfortable. Similarly, in the realm of conscientiousness, one cannot be completely successful at finishing tasks in time if one cannot plan ahead, and for finishing tasks, it generally helps if a person enjoys working hard. Some of these properties are connected, in the sense that they are mutually dependent on one another. Other properties are unconnected or very weakly connected (i.e. relatively mutually independent). For instance, one can like working hard without being able to make friends easily. Thus, these dependencies between personality components define the structure of the network that characterizes a person, that is, is the personality architecture (Cervone, 2005).
Now, suppose that one settles into a behavioural equilibrium with respect to one property. Say, a person likes to be around people and as a child seeks the company of others systematically (for a similar point of view, see Caspi, Bem & Elder, 1989; Caspi, Elder & Bem, 1987, 1988). As a result, one's social skills are developed and improve over time, which makes it easier to be around others, among others, until at some point an equilibrium is reached. This means that the situation has become relatively stable: one likes to be around others, and one has succeeded in finding a way to realize that state (e.g. a job in a social environment), barring situations where one is temporarily and involuntary ‘kicked out’ of equilibrium (e.g. being ill and therefore unable to leave the house for some time). Then, the evolution of this property (i.e. enjoying the company of others) will cause other properties, such as social skills, to co-evolve into a related equilibrium: it is quite hard to like to be around people and actively seek out environments that match this preference without at the same time developing social skills. Another person may reach the same equilibrium but approach it from the other direction; for some reason, the person becomes highly skilled in social interactions and comes to like the company of people as a result. Thus, groups of properties will move synchronously, like a flock of birds or a swarm of bees, simply because the organization of the human system and its environments require it.
This idea stands in stark contrast with the idea that behaviour is caused by a small set of latent personality dimensions/traits. In terms of the flock of birds analogy: in the situation as mentioned earlier, one bird in the flock flies in a particular direction because its neighbouring birds do so; in a latent trait scenario, all birds in the flock fly in a particular direction because of the instructions of an invisible (i.e. latent) bird. That is, in the standard model, personality dimensions/traits function as ‘common causes’ of a set of item responses (Borsboom, 2008a; Edwards & Bagozzi, 2000; Schmittmann et al., ). In psychometric terms, one of the most important features of a latent trait model that signals this assumption is ‘local independence’ (e.g. Holland & Rosenbaum, 1986; Lord, 1953; McDonald, 1981). Local independence means that, conditional on any given position on the latent variable, the observed item responses are statistically independent. Essentially, this means that the associations between items are spurious in the sense that they arise ‘solely’ from the items' common dependence on the latent variable. This is structurally analogous to the textbook example of the correlation between the number of storks and the number of newborns across Macedonian villages: villages that have more storks also have more newborns. This association is spurious because the correlation between storks and newborns arises solely from both variables' dependence on village size: larger villages have more chimneys, which attracts storks, and more people, who produce babies.
A latent variable model works in the same way. It relies on the assumption that dependencies among the cognitive, affective and behavioural components of personality (i.e. the individual birds in the flock, for example neuroticism items) arise ‘solely’ because all components depend on the same underlying trait (i.e. the invisible bird, for example neuroticism). Figure 1 shows an application of this model to the Big Five dimensions as measured with the NEO-PI in 500 first year psychology students at the University of Amsterdam (see Dolan, Oort, Stoel & Wicherts, 2009). Reliance on the assumption of local independence is evident by the absence of any direct connections between items. As such, local independence explicitly prohibits direct causal relations between the components of personality as represented by the items. The model with five latent traits influencing only their respective items—as depicted in Figure 1—does not fit the data (df = 28430, χ2 = 60839, p < .001), which is mainly due to violations of simple structure: particularly, the correlations between items that belong to different personality dimensions are too high to be accounted for by the model. How can one address this problem? One way, the standard way in personality psychology (e.g. Savla, Davey, Costa & Whitfield, 2007) is by tweaking the model ‘on the basis of the data’ so that the basic latent variable hypothesis is preserved (e.g. by allowing cross-loadings, exploratory factor analysis with procrustes rotation; see also Borsboom, 2006 for an elaborate critique). Another way would be to make the simple structure model more complex, for example, by introducing first-order and second-order latent variables (not to detract from the main aim of this paper, we have not included fitting such more complex models). Another way, the central tenet of this paper, is to consider the misfit of the un-tweaked model, an indication that the latent variable hypothesis fails as an explanation of the emergence of normal personality dimensions, and to move on towards alternative models.
That is, because of the local independence assumption, the very idea of cognitive, affective and behavioural components (i.e. items) that are directly connected to one another for causal or homeostatic reasons (or, for that matter, because of logical ones) is irreconcilable with the dominant latent trait perspective on personality dimensions and their items. If we take the connections between the components of personality to be real, that is, non-spurious, then a viable alternative approach is to describe them as a network. The crucial aspect of such a network is its organization: the way in which functional components of human personality are linked to one another. In turn, this organization depends critically on the equilibriums of the human system and its environments: certain behaviours correlate or coincide, whereas others do not because they are compatible or incompatible, respectively, with respect to specific equilibriums.
From this point of view, neuroticism items are tightly connected not because they are caused by the same latent trait (neuroticism) but because they arise in similar equilibriums: for example, someone who feels threatened easily will likely also suffer from nerves, feel lonely and worry too long after an embarrassing experience. Items related to the free exploration of environments (e.g. being open to new people) will unlikely co-evolve within threat-related equilibriums and hence will not be tightly connected to neuroticism items. This is not to say it is no longer valid to speak of ‘neuroticism’ or ‘openness’ as personality dimensions/traits: it certainly is, but under the assumption of a network perspective, these terms do not indicate latent causes of behaviour but groups of tightly inter-connected personality components. Thus, we can still use a term such as ‘neuroticism’ to refer to a phenomenon that emerges as a result of the biological, psychological and environmental forces that knit some behaviours closely together. However, we speak of such a phenomenon just like we speak of a flock of birds. We know that a flock emerges out of the synchronized behaviour of the birds it contains and would not venture to hypothesize that it existed independently of that behaviour, let alone was caused by it (Schmittmann et al., ).
Naturally, we are not the first to raise questions about the incompatibility of current trait models with dynamic interactions between personality components and the environment. Similar ideas have been manifested in the writings of personality theorists, almost since the inception of the discipline; recent theorists such as Mischel & Shoda (1995) and Cervone (2005), as well as Read et al. (2010), have argued along very similar lines. However, the methodology to study complex networks has been developed to maturity only relatively recently (e.g. Albert & Barabási, 1999; Newman, 2006; Watts & Strogatz, 1998). As a result, we are now able to use such techniques to visualize and analyze large-scale networks in ways that have not been possible before. The remainder of this paper aims to give first passes at applying these ideas systematically in the study of personality. We focus on four illustrations regarding (i) the overall organization of personality components, (ii) the distinction between state and trait, (iii) the genetic architecture of personality and (iv) the relation between personality and psychopathology.
A NETWORK OF PERSONALITY COMPONENTS
Mapping the structure of personality onto a network is a daunting task. Fortunately, we have a reasonable starting point in the form of common personality questionnaires that query respondents for their status with respect to exactly the type of components that would be likely candidates to make up a personality network structure. The correlations between components will tend to be higher when the connectivity in the human system is stronger. Thus, by studying correlations and representing them in a network structure, one may obtain a first glance at the visualization of the global (i.e. average) structure of personality components. We have developed an R-package for network analysis (qgraph: Epskamp, Cramer, Waldorp, Schmittmann & Borsboom, 2011) that is capable of constructing such visualizations directly from the data. In essence, the routines in this package treat a correlation matrix as a so-called weighted network, that is, it treats the items as components and their correlations as the strength of the connections among these components. The result of applying this routine to the items of the NEO-PI-R is represented in Figure 2 (for the large central graph, same sample as used for Figure 1; for the small graph in the top right, simulated data).2
A graph like that in Figure 2 offers a powerful visualization that can be used to reveal patterns and structures that would be very difficult to spot by using traditional methodology (note that Figure 2 represents the complex structure of no less than 240 × 240 = 57600 correlations with little data reduction). For instance, looking at Figure 2, there are a few things that catch the eye immediately. First, the network is very densely connected, much more connected than would be expected if a small number of latent variables gave rise to the correlational structure (even if we let these five latent variables correlate, as we did in Figure 1). In particular, this visualized pattern of correlations between personality items is not convincingly suggestive of five distinct latent traits. This can also be seen when comparing the empirically constructed graph with the inserted graph at the top right of the figure, which shows the correlations that would be expected if the five-factor model of personality were true (i.e. if the covariation between items could be solely explained by five correlated latent variables that cause the item responses).
In this dataset, the strongest organization arises for neuroticism and conscientiousness items (red and purple nodes/circles in Figure 2). Extraversion and agreeableness items (yellow and blue nodes in Figure 2) are largely intertwined with one another, meaning that, on average, an extraversion item is not much more strongly correlated with other extraversion items than with agreeableness items (and vice versa; difference in average correlations is 0.12). This makes sense from a network perspective. For instance, it becomes easier to spend time with others (agreeableness) if one likes to be around others (extraversion), and it is difficult to talk much with people at parties (extraversion) when one is not really interested in others (reversed agreeableness item).
Another interesting aspect of the graph in Figure 2 is that some items are more strongly connected to other items (those items are placed towards the middle of the graph: e.g. nodes representing item numbers 15, 48, 49, 135 and 229), whereas others are only weakly connected to other items or not connected at all (those items are placed towards the periphery of the graph: e.g. nodes 88 and 239). That is, some items are more ‘central’ in the network than others (see also Cramer et al., 2010). Without a network representation, one would be very unlikely even to think about a concept such as centrality in this way, let alone think of ways of computing it.
For example, the item ‘When I promise something, one can count on me to fulfill that promise’ (node 135) is a central item in the Big Five network. This makes sense because the content of that item is closely connected to not only other conscientiousness items—for example, to fulfil a promise, one generally has to be a reliable person (node 45) and have a tendency to finish things one has started (node 145)—but also items of other personality dimensions as well (i.e. thick lines in Figure 2): for example, someone who likes people and sympathizes with them is more likely to fulfil a promise (agreeableness, node 126) as well as be someone to whom others turn when decisions have to be made (extraversion, node 132). On the contrary, the item ‘We can never do too much for the poor and the elderly’ (node 89) is a peripheral item: other than a few connections with other agreeableness items—people who care about the poor and the elderly generally feel sympathetic towards people who are worse off (node 209)—(not) caring about the poor and the elderly has (very) little to do with how open, extraverted, neurotic and/or conscientious one is. Thus, items in the Big Five network differ in terms of their centrality in that network, and given the content of the items, these differences in centrality appear to make theoretical sense.
Importantly, the entire notion of central versus peripheral components in the Big Five network is irreconcilable with a latent trait perspective on personality, which is articulated in a latent variable ‘measurement’ model: in such a model, save for measurement error, items that measure the same trait are exchangeable and thus equally central or peripheral (factor loadings are reliability estimates and as such, cannot be measures of centrality as we view the concept). For instance, if the latent variable model in Figure 1 were true, then someone's position on the conscientiousness continuum could be determined perfectly from knowing the expected value of any one of the conscientiousness items (Jöreskog, 1971; Lord & Novick, 1968; see also Borsboom, 2005). In other words, the model holds that if one knew the expected value of a person, say, on the item ‘I tend to finish things once started’, then none of the other items would offer any additional information about how conscientious that person is (i.e. that person's position on the latent conscientiousness continuum). In that sense, all items are equally central (or peripheral), just like mercury thermometers are no more ‘central to temperature’ than digital or other thermometers are.
Does it matter if some components in the Big Five network are more central than others? It does because, first, it hints at which pathways are more likely to result in the emergence of certain personality structures in some people. A person's personality structure can be represented in a network analogous to the one in Figure 2 (for such an individual network, connection strength then refers to how strongly two personality components are connected over time in one person), a subject we will return to in the next chapter. Because Figure 2 is based on between-subjects data (and is, as such, an ‘aggregation’ of the networks of all these individual subjects), it is likely that at least in some of these subjects, the central components in Figure 2 are prominent features in their networks as well. The network model predicts that once such a central component becomes ‘active’ in someone (i.e. a component changes in terms of its state,3 for example, not having experienced this before, someone starts to experience fear of disappointing others, a component linked to both agreeableness and extraversion, Mongrain, 1993), then the probability of neighbouring components to become active as well rises because of the strong connections of that component with other components in the network (e.g. ‘I get chores done right away’ and ‘I finish things I have started’). This particular pathway (fear of disappointing others ↔ getting chores done ↔ finishing things) to a personality structure in which multiple conscientiousness items are active is then more likely than a pathway to conscientiousness that includes peripheral components (e.g. ‘I take voting and other duties as a citizen very seriously’, node 35).
Second, centrality matters because it is linked to the ability to change and to how widely spread out the consequences of such change will be. When a personality component is central, it is likely to be dependent on many other components (and vice versa), so it will be more difficult to change. Changing a habit of not fulfilling promises, for example, is more likely to be difficult because to change that component, there are many others that may need to be changed as well (e.g. sympathize more with other people's needs and learning to finish things). Drawing analogy to a trade network, there are tradesmen who operate as pivotal points in the network (i.e. as central nodes): they have a large influence on the total productivity of the network (how much money the network as a whole makes), and it is very difficult to drive them out of business because of their strong connections with so many others (i.e. individual components of personality). It is unlikely but for some reason, it might be that a central component in fact does change (in the trade network analogy, a pivotal person goes out of business). If so, then the consequences for the remainder of the network will be more widespread than if a peripheral component (tradesman on the periphery) changes. For instance, if one ceases to take voting seriously, this is not likely to have major effects on other aspects of one's expression of personality. In contrast, if because of whatever circumstance, one ceases to be a reliable person—as might occur in the early phases of dementia with a deterioration of memory functions—this is likely to have effects throughout the system.
NETWORK STRUCTURE AS A SOURCE OF STABILITY
Human actions are flexible and unpredictable across situations, but at the same time, general patterns of behaviour can be extremely rigid and very difficult to change. Theories of personality aim to reconcile these two facts of human life and provide compelling explanations for the stability that apparently underlies the great variability in daily moods, thoughts and behaviours. The traditional way of dealing with this issue is to invoke a two-part explanation in which the variation in behaviour is governed by transient factors, whereas the average around which these variations are dispersed is caused by a stable factor. The latter is typically conceptualized as a trait, defined as a relatively enduring organismic (psychological, psychobiological) structure underlying an extended family of behavioural dispositions (Tellegen, 1991). Thus, in this definition, a ‘trait’ is a ‘common cause’, a structure that ‘explains’ the stable level of functioning around which a certain variability in ‘states’ revolves (e.g. the trait-state-error model; see Kenny & Zautra, 1995). An example of such a structure is the latent dimension of extraversion, which is thought to cause stability by affecting the chances for a broad range of states to occur, as shown in the left panel of Figure 3.
Both the traditional model in the left panel of Figure 3 and the network in Figure 2 are between-subjects models that—for several reasons—cannot be assumed automatically to generalize to specific individuals (Borsboom, Mellenbergh & van Heerden, 2003; Borsboom, Kievit, Cervone & Hood, 2009). From a network perspective, inference at the level of the individual is possible if one assumes that the dynamic structure of personality components of an individual can be represented in a similar network form (e.g. Figure 4 is an example of such a hypothetical network of an individual). Individual differences can then be captured by allowing for individual differences in components and the strengths of the connections among them.
From a network perspective, there are multiple ways in which trait-like and state-like characteristics can be defined at the level of individual networks (see Figure 4 for an illustration). This flexibility stands in stark contrast to the trait view, in which traits and states can only be sensibly defined at the level of the (first-order or second-order) latent dimensions (from a latent trait perspective, it would make no sense to define states and traits at the level of the items, although it might be technically possible). As such, individual differences can only be expressed in terms of that latent dimension as in, for example, ‘Alice is more trait neurotic than Bob’, whereas the network perspective can express many differences between Alice and Bob, such as ‘Alice's network has more trait-like neurotic components than Bob's’ and ‘The connections in Alice's network are more state-like than Bob's’. Such multiple observations are likely more true to the subtle nature of individual differences.
A first way to define traits and states in a network is at the level of the network as a whole (turquoise circle around the entire network in Figure 4): synchronized stability of multiple components can result in the emergence of a stable trait such as extraversion (as illustrated in the right panel of Figure 3, in this case for extraversion components). That is, instead of the current view that a trait ultimately ‘causes’ behaviour (i.e. arrows pointing from Extraversion in the left panel of Figure 3), the network perspective views a trait as a phenomenon that is the ‘result’ of (and, in that sense, emerging from) direct interactions between behaviours as measured with personality items (i.e. arrows pointing towards Extraversion in the right panel of Figure 3). As such, a trait, from a network perspective, is similar to a summary statistic or index variable that describes the average activation level of states, which is consistent with the key assumption of a formative model (see Edwards & Bagozzi, 2000). Importantly, the network perspective is thus ‘not’ contradictory to current trait theories. Networks result in traits too, the only difference with current trait theories being that in the latter case, traits are most typically explained with a latent variable model in mind. That is, trait theories are currently intertwined with a latent variable perspective (as depicted in the left panel of Figure 3).
Networks can result in traits because transient factors and context determine the activation of affective, cognitive and behavioural components that in turn may activate one another if they are connected in the network architecture. Every time a set of components is activated (e.g. when a person feels energetic), the activation contributes to a self-evaluation stored in memory (‘I am an energetic person’) that is of the kind queried in typical personality questionnaires (‘Would you consider yourself an energetic person?’) and serves as evidence for evaluating the self-related hypothesis (van der Maas, Molenaar, Maris, Kievit, & Borsboom, 2011). General evaluations that arise from densely connected areas in the network will covary; as a result, these variables will form a large principal component if submitted to a data reduction technique such as principal components analysis. However, a simple structure confirmatory factor model (like that in the left panel of Figure 3) may not fit well because of violations of conditional independence (i.e. because the model does not get the causal structure right). We understand this to be typical in personality research where confirmatory models can fit badly even though principal component structures are robust and replicable (McCrae, Zonderman, Costa, Bond & Paunonen, 1996).
A second way in which networks can display trait-like and state-like properties in individuals' networks is at the level of these individual components themselves (turquoise and pink circles around the item boxes in Figure 4). Roughly speaking, there are two reasons why components can display both trait-like and/or state-like features. First, the wordings of the items themselves may or may not refer to stable behavioural dispositions. For example, the Big Five network presented earlier (Figure 2) contains many components that refer to stable behavioural dispositions (e.g. ‘I easily feel offended by other people’ and ‘I finish things once I have started them’), whereas the responses to other items may greatly vary over time (e.g. ‘I feel offended now’ or the items represented in Figure 5). The latter components can be considered to be inherently more state-like, whereas the first are inherently more trait-like. Second, the activation of components can be altered (from ‘active’ to ‘not active’ or vice versa), depending on a specific ‘situation’ a person is in (orange arrow from situation to i2 in Figure 4). Some of these components are more state-like because alterations in the environment (i.e. different situations) result in unstable activity patterns (i.e. the change in activity is relatively temporary). For example, a component such as ‘I'm full of ideas’ can be unstable in certain people: the component would be active (i.e. Alice feels full of ideas) for Alice after a positive day at work during which her boss complimented her on having a good idea but inactive (i.e. Alice does not feel full of ideas) the next day because her mother-in-law describes her, in her face, as a follower and not a leader. In contrast, some situations result in long-term stable changed activity in one or more nodes. For example, Bob, a trusting person, obtains a venereal disease from his cheating girlfriend who also dumps him. Subsequently, Bob re-examines basic assumptions about how he sees the world and as a result, changes: becomes less trusting, more suspicious of the motivations of others and so on.
Situations can also influence the connections among the components (orange arrow from situation to the connection between i1 and i2 in Figure 4; analogous to moderation). Connections subject to such influences can be more susceptible to change and thus more state-like in that they are aspects of personality that vary in response to different situations (analogous to what is hypothesized in the CAPS model: Mischel & Shoda, 1995, 1998). For example, Bob normally does not feel guilty because he sometimes feels just miserable for no reason (i.e. relatively stable weak connection between feeling miserable and feeling guilty). But, when Bob feels just miserable right when his wife surprises him with tickets for a cruise, he feels incredibly guilty: that is, the connection between feeling miserable and feeling guilty is stronger, triggered by the situation. It is, on a related note, this very malleability of certain connections that is the focus of many psychological treatment strategies (e.g. cognitive behavioural therapy; see Cramer et al., 2010). Other connections are likely relatively trait-like, in part, because the components they connect are inherently more trait-like as well, for example, the connection between ‘I like to go to parties’ and ‘I feel comfortable around people’.
The empirical study of this dynamic structure of personality networks becomes possible through the use of time-series data. For instance, Figure 5 presents empirical correlation networks of four people who participated in a larger study into the effects of mindfulness training on a range of emotion and psychopathology variables (Geschwind, Peeters, Drukker, van Os & Wichers, 2011; see Appendix A for a description of the sample and the measures). The participants in this study were assessed multiple times a day by using an experience sampling protocol, which generates series of observations over time. The networks in Figure 5 represent the lag-1 correlations between time series of four variables: anxiety, feeling down, irritability and the pleasantness of the event reported to be the most important one during the assessment period. Specifically, a thick green arrow from A to I means that a higher level of anxiety at t predicts a higher level of irritability at t + 1, a thick red arrow from E to A means that a more positive evaluation of the event that took place at t predicts a lower score on anxiety at t + 1 and so on. Naturally, it is also possible to construct such intra-individual networks for correlations within the same time frame: the construction and interpretation of such graphs would be analogous to the procedure explicated for the inter-individual network that was presented in Figure 2.
The individuals showed marked differences in their dynamic structure. Individual 1, who is relatively typical for the sample studied here, showed positive dependencies among A, I and D: for example, the more anxious at t and the more irritable at t + 1 (and vice versa). A, I and D all had negative dependencies with E: for example, lower anxiety at t predicted higher pleasantness of the event reported at t + 1, but a more pleasant event at t also predicted lower anxiety at t + 1. This appears not to be the case for individual 2 whose relations between E and the psychological variables were one-way traffic: for example, lower anxiety at t predicted higher pleasantness of the event at t + 1, but a more pleasant event at t did not appear to predict lower anxiety at t + 1. One might speculate that this individual ‘profits’ less from positive events. Participant 3 also showed this pattern but in addition showed no noticeable predictive relation between the psychological variables at t and the pleasantness of the event at t + 1. This individual thus appeared to function independently of the events reported in the relevant time. Finally, participant 4 showed a surprising pattern of purely negative relations, in which the anxiety variable functioned as a source node without substantial incoming effects and seems to steer the other variables in a counterintuitive way (‘increased’ anxiety at t predicted ‘decreased’ irritability and depressed mood at t + 1, whereas ‘decreased’ irritability and depressed mood at t predicted a ‘more’ pleasant event at t + 1). We do not know, from the present data, to what extent these patterns generalize outside the studied time window or whether they have meaningful connections to the everyday functioning of the studied individuals. However, the differences between the network structures are quite suggestive and may, in future research, be shown to have significant consequences.
Thus, from a network perspective, the components of individuals' personality networks as well as the connections among them can exhibit trait and/or state-like properties, in part influenced by situations that figure as separate nodes in the network (see also Figure 4), and traits such as extraversion, or openness, emerge out of the combined activity of the components of the personality network, instead of being the common cause of these components. Traits as emerging entities are not in violation of some definitions of traits: for instance, the definition of Tellegen (1991) of traits as ‘enduring […] structure[s] underlying an extended family of behavioural dispositions’ would in fact seem neutral on whether the structure in question is a latent dimension or a network structure.
Understood in this way, the network perspective offers a possible resolution between trait approaches and situationist approaches that emphasize that traits can be adequately described as situation-relevant reaction patterns (e.g. Mischel & Shoda, 1995): the connections among situational nodes—external to the human system—and components that are more internal to the human system are likely to differ in strength across individuals. Such differences in situation → behaviour associations lead to if-then signatures of the kind identified by Mischel and Shoda (1995).
Given the ample opportunities for individual differences to arise in a personality network structure, is it in fact possible that ‘both’ stable individual differences ‘and’ significant day-to-day variation arise from the same network structure? The answer is yes. To illustrate this, Figure 6 shows three simple networks, representing three fictitious people, consisting of three binary nodes (i.e. nodes that can be either ‘active’: 1 or ‘inactive’: 0). All variables are ‘measured’ at multiple time points but without implying any direction of causation (i.e. all variables influence all other variables). Thus, each resulting network is an intra-individual representation of how three variables influence one another bidirectionally over time. The only differences among these networks are the strengths of the connections among the nodes (i.e. each connection has a certain weight that determines its strength): the rightmost network in Figure 6 is the most strongly connected, whereas the leftmost network is the least strongly connected. At time point t, whether or not a node is active is dependent on the status (0 or 1) of each of its neighbours times the relevant connection weight, which results in a total incoming effect A. The probability that the node is active then depends on the total incoming activation as follows: P(node activet+1) = 1/(1 + eAt) (note the similarity of this equation to equations in item response theory; e.g. Lord, 1953).
If we simulate data points according to this model for the three networks in Figure 6, the networks will all transition between activation patterns in a random fashion. That is, there will be significant (‘day-to-day’) variation in which nodes are active or inactive. On the other hand, the probability distributions of the total activation scores (i.e. the total number of nodes that are active at a randomly chosen time point for each network) will be stable: in Figure 6, the average activation level of the leftmost network will be lowest, whereas that of the rightmost network will be the highest, and this is no surprise, given the fact that those networks are weakly and strongly connected, respectively. Thus, stable individual differences in average activation levels are possible as well, and it is exactly that synchronized activity of consistent patterns of node activation within individuals may give rise to traits: if Figure 6 represented openness networks, then the person with the rightmost network would likely be an open person—because, on average, many openness nodes are active at the same time—whereas the person with the leftmost network would likely not be an open person. So given this potential of individual differences in network structure to generate both traits and day-to-day variations without invoking any latent dimensions, what, in turn, could cause these differences in network structure?
THE GENETIC ORIGINS OF PERSONALITY NETWORKS
Genes influence many human characteristics, and personality is one of them. Multiple studies have shown that personality dimensions are at least moderately heritable (Boomsma, Busjahn & Peltonen, 2002; Bouchard, 1994; Jang, Livesley & Vernon, 1996; Kendler & Myers, 2010; Riemann, Angleitner & Strelau, 1997; Loehlin, 1992): for example, 40% of the phenotypic variance in extraversion can be explained by additive genetic factors.
Assigning one number to represent heritability of any particular personality dimension makes sense from a latent trait perspective: items are no more than indicators of a common underlying trait, say, extraversion, and as such, what is transmitted via genes from one generation to the next is the predisposition for developing that personality trait not the propensity for a particular type of behaviour/emotion/cognition as measured with a single item (i.e. personality component; see the left panel of Figure 7). In analogy with height, height is a latent trait (i.e. height is an unobserved variable for which we need measurement instruments to quantify it in individuals; see Bollen, 2002; Borsboom, 2008b),4 which is measured with various methods (e.g. measurement tape). What is heritable is relative height itself—that is, children of tall parents tend to be tall as well—not any particular measurement of height.
One can go one step beyond defining heritability as a characteristic of a personality dimension such as extraversion (whether it be a latent trait or an emerging feature, as it is defined in a network model), namely, by defining heritability as genetic influence on the individual components of the network and the connections between these components (see the right panel of Figure 7). For example, it might be that liking parties is 20% heritable, whereas enjoying the company of other people is 65% heritable, the difference not being due to differences in reliability (consistent with a latent independent pathways model). Likewise, it could make sense to say that the degree to which people who are quick to understand things have a tendency to be full of ideas (i.e. connection strength between these two components) is 34% heritable or the degree to which people who regularly feel just miserable have a tendency for suicidal ideation (and vice versa) is 78% heritable.
Now, if the network perspective is accurate in portraying personality, then current techniques for the next step in behavioural genetic research, that is, the identification of genes that are the driving forces behind these heritability estimates, might prove problematic. Current techniques employed in genetic association studies typically rely on a sum score (e.g. the sum of the neuroticism item scores of the NEO-PI) as a proxy for the latent variable, neuroticism in this example. Genetic association studies in their most rudimentary form identify genes or genetic variants as being associated with a particular personality dimension if they predict the sum score (i.e. the dependent variable in the design; Cramer, Kendler & Borsboom, 2011; van der Sluis, Kan & Dolan, 2010). If personality dimensions were indeed latent traits, this approach is sensible, although not necessarily optimal (van der Sluis, Verhage, Posthuma & Dolan, 2010).
To date, standard genetic linkage and association studies have not yielded any clear genetic candidates: for the Big Five personality dimensions, many candidate gene findings are not replicated, and the genetic polymorphisms that are consistently identified typically account for less than 2% of the genetic variance (Amin et al., 2011; de Moor et al., 2011; Fullerton et al., 2003; Kuo et al., 2007; Nash et al., 2004; Terracciano et al., 2010). This discrepancy between moderately high estimates of population heritability and the inability to identify the responsible genetic polymorphisms is called the ‘missing heritability’ problem, a problem that is pervasive throughout the entire realm of psychology as well as other complex biomedical traits such as height and blood pressure (e.g. Maher, 2008; Manolio et al., 2009).
Although many explanations have been put forward for this missing heritability problem (e.g. additive small effects of many individual genes, limited sample size, population stratification and selection bias; Frazer, Murray, Schork & Topo, 2009; Maher, 2008; Sullivan, 2011), we focus on another possible reason: misconceptualization of the phenotypic model (Figure 7). In particular, the model in the left panel of Figure 7 might be wrong (as was, for example, recently shown for nicotine dependence where two genes influenced individual symptoms quite differently: Maes et al., 2011). From a network perspective (the right panel of Figure 7), it is not likely that all components and connections between them in the personality network are influenced by the exact same set of genes (see the right panel of Figure 7: gene 1 influences other parts of the network than gene 2). For example, components such as feeling sad and finding political discussions boring probably involve different antecedent pathways: feeling sad has more to do with emotional processes, whereas finding political discussions boring is more likely a cognitive phenomenon, and as such, feeling sad and finding political discussions boring probably involve different biological substrates and pathways and thus different genes. If so, then attempting to relate genetic polymorphisms to their sum score is not likely to contribute to effective gene hunting because with a sum score, one only captures the genetic variance that is shared among the components and their connections (van der Sluis et al., 2010): the power to detect effects from single-nucleotide polymorphisms (SNPs) in sum scores is multiple times lower when these gene effects are local (e.g. gene 2 in the right panel of Figure 7 influences two neuroticism items) compared with when these effects are global (e.g. gene 1 in the left panel of Figure 7 influences all neuroticism items via the latent trait ‘neuroticism’).
It is hard to pit the models in Figure 7 directly against each other because estimation and fitting algorithms for the network model have not been developed in sufficient detail. However, we can examine and test divergent predictions of the models such as the location of the effect of SNPs: from a latent trait perspective, one would expect SNPs impact at the latent trait level, whereas from a network perspective, one expects SNPs to impact at the level of the individual components (see Maes et al., 2011).
We tested this prediction by using data from 1625 healthy individuals who participated in the dbGAP GAIN Major Depression Disorder study (dbGAP study accession, phs000020.v2.p1). In particular, we investigated the effects of seven top SNPs that were implicated in neuroticism in two recent genome-wide association studies (de Moor et al., 2011, Terracciano et al., 2010). We tested whether the effects of these genes on the item responses were most likely to be mediated by the latent trait ‘neuroticism’ or whether these effects were more likely to be item specific (see Appendix B for an extended description of the sample and the method).
The analyses showed that in this sample, none of the seven top SNPs had a significant direct influence on the latent trait ‘neuroticism’. On the one hand, this result can be interpreted as a non-replication of these SNPs in this sample, which could be due to a limited sample size or the use of a different neuroticism instrument (see Appendix B). On the other hand, the result can be interpreted as lack of support for a latent trait perspective on the influence of genes on personality dimensions. At the same time and in support of the network perspective, we did find evidence for significant direct influences of three SNPs (rs17453815, rs12509930 and rs7329003) on three individual neuroticism items (‘restless, can't sit still’, ‘guilty’ and ‘sleepless due to thought racing’: see Figure 8). These effects were significant at α = .005 (p-values for the SNP-latent trait ‘neuroticism’ ranged between .16 and .62; see Appendix B). Naturally, replication of these specific SNP-item relations in other, larger samples is imperative to draw definitive conclusions. This example with real data mainly serves to illustrate how to test the diverging predictions from the latent variable versus the network perspective.
Another way of testing the viability of the latent trait perspective is to check whether the directions of the effects of the seven top SNPs (i.e. increase or decrease risk) are the same across the individual neuroticism items. If the latent trait perspective is correct and genes influence individual items only indirectly via the latent trait ‘neuroticism’, then all relations between an SNP and identically coded neuroticism items should have the same sign. This, however, is not what we found in this data set (see Table S1 in Appendix B for the odds ratios (OR) between SNPs and neuroticism items and p-values computed according to false discovery rate criteria). For example, SNP rs17453815 was associated with a decreased risk for being ‘easily irritable’ (OR = 0.69; .01 < p < .05) but with an increased risk for ‘restless, can't sit still for long’ (OR = 1.32; .001 < p < .01). Similarly, SNP rs11707952 was associated with a decreased risk for ‘experiencing mood swings’ (OR = 0.73; .01 < p < .05) but also with an increased risk for ‘not feeling your old self’ (OR = 1.22; .01 < p < .05).
Given that the current methodological state of affairs does not allow for a direct statistical test, psychometric modelling with genetic data might provide a fruitful avenue to explore the feasibility of latent trait versus network models because these models come with specific predictions that can be tested in a confirmatory factor analytic framework. As such, we do not take the aforementioned results to signify anything definitive about SNP effects on neuroticism items; rather, these results serve as concrete examples of how one might go about testing predictions of the two competing models. It might be argued that finding local effects (i.e. SNP effects on individual items) is not in violation of the ‘statistical’ aspects of the latent trait model. Although this is true, the ‘theoretical’ notion of a latent variable as an accurate reflection of personality dimensions is much harder to maintain in the face of genetic effects whose impact is not at that latent level but, instead, at the item level.
If future evidence favours the network model, the next step would be to wonder how personality networks are tied to psychopathological phenomena. As we have argued in earlier work (Borsboom, 2008a; Cramer, Borsboom, Aggen & Kendler, 2011; Cramer et al., 2011; Cramer et al., 2010), mental disorders can also be understood in terms of networks of interacting ‘symptoms’ (e.g. insomnia → fatigue → concentration problems). Because it is well known that certain personality dimensions predict the development of certain forms of psychopathology (e.g. Hettema, Neale, Myers, Prescott & Kendler, 2006; Kendler, Gatz, Gardner & Pedersen, 2006; Terracciano, Lockenhoff, Crum, Bienvenu & Costa, 2008; van Os & Jones, 2001), how might this covariation arise from a network perspective?
THE ROADS FROM PERSONALITY DIMENSIONS TO MENTAL DISORDERS
Some aspects of personality are correlated with the onset and/or maintenance of certain mental disorders: for example, (i) trait neuroticism and major depression (MD), (ii) alienation (a tendency to feel mistreated, victimized, betrayed and the target of false rumours) and substance dependence and (iii) high negative emotionality (a propensity to experience aversive affective states) and antisocial personality disorder (e.g. Klein, Kotov & Bufferd, 2011; Krueger, 1999; Krueger, Caspi, Moffitt, Silva & McGee, 1996). From a latent variable perspective—in which a personality dimension and a mental disorder are latent entities—there are three ways in which personality features (P) and mental disorders (M) can be modelled (see Figure 9): (i) models in which P and M are not causally related in whatever shape or form. Instead, P and M are correlated because they are (partly) influenced by the same etiological processes (the A arrows in Figure 9); (ii) models in which P is an effect of M (the B arrow in Figure 9); and (iii) models in which P precedes M (the C arrow in Figure 9).
From a network perspective, the three classes of models as depicted in Figure 9 do not work because in both personality and mental disorder networks, there are no latent variables. Because items are at the heart of personality networks and symptoms at the heart of mental disorder networks, the most sensible way to conceive of relations between the two networks is by means of direct relations between these items and symptoms (see blue lines in Figure 10). Instead of one option for three types of pathways between personality and psychopathology (i.e. A, B and C model in Figure 9), each blue line between an item and a symptom in Figure 10 represents an optional pathway that could be of the A, B or C type. For example, in Figure 10 (without implying causality because there are no arrows in the figure), one pathway from personality to mental disorder (and vice versa) could be: p2–p4–m5 or, alternatively, m2–p3–p5. That is, from a network perspective, pathways between items and symptoms indicate dependencies between them, such that one may activate another, analogous to how diseases spread through a population. For example, the tendency to feel nervous around other people (p2) likely increases the probability of spending much time alone (p4), which may result in relatively frequent feelings of anhedonia (m5). The other way around may be an equally likely pathway: prolonged feelings of anhedonia may well undermine the capacity to enjoy the company of other people.
As a starting point, like in Figure 2, correlations between personality items and mental disorder symptoms could be used as quantifications of the strength of the connections between these items and symptoms. Figure 11 shows such a correlation network for neuroticism and MD data obtained from the Virginia Adult Twin Study of Psychiatric and Substance Use Disorders (Kendler & Prescott, 2006; Prescott, Aggen & Kendler, 2000; see Appendix C for a description of the sample and the measures). Some marked differences in connection strengths among the items and symptoms stand out. First, there are clearly two clusters of strongly connected items/symptoms, one corresponding to neuroticism (blue nodes) and the other corresponding to MD (red nodes). Second, some neuroticism items are more strongly connected to MD symptoms than other neuroticism items (and vice versa): for example, feelings of worthlessness (wort: MD) and feelings of loneliness (lone: neuroticism) are more strongly connected than one's feelings being easily hurt (hurt: neuroticism) and increased appetite (iapp: MD).
Within a network such as the one in Figure 11, central nodes might be the crucial nodes on pathways connecting neuroticism and MD because such nodes are strongly connected with both neuroticism and MD nodes (as argued earlier in the paper). So for example, in this particular sample, feeling just miserable (mise: neuroticism), being a nervous person (nerv: neuroticism), feelings of loneliness (lone: neuroticism) and feelings of worthlessness (wort: MD) are the most likely candidates for being part of the multiple pathways from neuroticism to MD.
Another way of generating hypotheses about likely pathways from personality to psychopathology (and vice versa) is through ‘partial’ correlations. The general idea is the same as with simple correlations—one constructs a network with the strengths of the connections between the nodes reflecting the magnitude of the correlations—but partial correlations are potentially more informative about whether two variables are in fact truly related. A high simple correlation between two variables does not necessarily imply that a unique relation exists between these variables. For instance, a high correlation between feelings of guilt and feelings of worthlessness may be due to the fact that both components are influenced by another component in the network, for example, depressed mood. As such, feelings of guilt and feelings of worthlessness are not uniquely related; the correlation arises because of their common cause, depressed mood. If that is true, the correlation between feelings of guilt and feelings of worthlessness should be (very) low when depressed mood is controlled, and this is exactly what a partial correlation does: it quantifies the association between any two components while controlling for one or multiple other components in the network. As such, when one computes correlations among the neuroticism and MD items/symptoms while controlling for ‘all’ other components in the network, a high partial correlation is potentially more indicative of a true relation than a simple correlation. Figure 12 presents such a partial correlation network on the basis of the same data that was used for Figure 11.
A few things stand out when inspecting Figure 12. First, many connections are markedly weaker in Figure 12 compared with connections between the same components in Figure 11, for example, the connection between feelings of worthlessness (wort: MD) and feelings of loneliness (lone: neuroticism): a direct relation between these components might exist (the partial correlation in Figure 12 is not close to 0) but is likely partially influenced by other components in the network (because the partial correlation is lower than the simple correlation). On the other hand, feelings of worthlessness (wort: MD) and feelings of guilt (guil: neuroticism) are almost as strongly connected in Figure 11 as in Figure 12: these two components are likely directly related without being substantially influenced by other components in the network. Second, some pathways from neuroticism to MD (and vice versa) are more likely than others (i.e. are more strongly connected compared with other pathways): for example, a pathway via feelings of worthlessness (wort: MD) and guilt (guilt: neuroticism) is more likely than a pathway via weight loss (wlos: MD) and describing oneself as a nervous person (nerp: neuroticism). When considering which nodes are the most central in this network, the most likely candidates for playing pivotal roles in pathways from neuroticism to MD (and vice versa) are feelings of loneliness (lone: neuroticism), guilt (guil: neuroticism) and worthlessness (wort: MD); thoughts of death (deat: MD), being nervous (nerv: neuroticism) and describing oneself as a nervous person (nerp: neuroticism).
Partial correlations may be used to generate more parsimoneous hypotheses about likely pathways from certain personality dimensions to certain mental disorders (and vice versa), but the technique is by no means bulletproof. It could be, for example, that a connection between two components with a low partial correlation does in fact exist. Sampling error, for example, might result in a low partial correlation between two components, whereas a direct relation in fact exists in the whole population. Therefore, replication of findings in multiple samples is a necessity before any definitive conclusion can be drawn. Another way of testing hypotheses generated by partial correlations in between-subjects data is via longitudinal studies in which these hypotheses are verified in individuals. But instead of focusing on total scores on personality and psychopathology questionnaires—which is the sensible thing to do when a unidimensional latent variable model holds (see Grayson, 19885)—longitudinal studies from a network perspective would analyze each item and symptom separately for a prolonged time. In addition, with the time-series techniques explicated earlier in this paper, the temporal pathways among these items and symptoms for individual people may be identified and directly modelled. Such studies undoubtedly will reveal many idiosyncracies—that is, there are likely many ways by which people develop certain forms of psychopathology as results of certain personality characteristics (and vice versa)—but with some strong between-persons partial correlations, we have found in the data example earlier that some important commonalities can be expected as well. As such, the network perspective and its associated investigation techniques may shed light on the exact nature of the complex relation between personality dimensions and mental disorders.
In the present paper, we have argued for a novel perspective on personality, in which the cognitive, affective and behavioural components of personality (e.g. liking parties and finding political discussions boring) are related through causal, homeostatic and logical connections. Traits such as extraversion and agreeableness emerge out of these connectivity structures, which implies a radical departure from traditional perspectives in which traits are causes of the relevant components. We have shown how the network perspective may potentially alter our conception of what personality is and may supply new research techniques to investigate (i) overall personality architecture, (ii) state and trait conceptualizations of personality, (iii) the genetic background of personality architecture and (iv) the relations between personality and psychopathology.
Naturally, network methodology is far from fully developed. Examples concern the development of estimation and fitting algorithms for network models, robustness analyses for inferences on network structures, combining inter-individual and intra-individual data and the question of model testing. Pertaining to the latter example, falsifying or confirming a network model can sometimes be quite complicated—for example, a unidimensional latent variable model will fit data that is generated by a network model in which all nodes are bidirectionally connected with equal strength—and sometimes surprisingly easy: for example, if one has the hypothesis that an inter-individual network is mutualistic (i.e. has only positive bidirectional connections so that nodes reinforce one another), then observing a negative correlation is enough to falsify that hypothesis. In its current state, it could be compared with latent variable modelling in the 1950s: we have the ideas and the models, but we still need to overcome many methodological obstacles. Nevertheless, the network perspective offers a plausible candidate model for explaining the ‘common’ structures of personality and the many idiosyncratic ways in which people deviate from that structure. One of its more attractive features is that the network perspective provides an intermediate position between traditional trait and situationist approaches, which both have longstanding traditions in personality psychology and which both have contributed greatly to our current understanding of personality. The network perspective takes the best of both worlds: it can explain how traits emerge out of the network structure, but it can also accommodate situational influences as external nodes that can activate individual components of the network (or connections among them).
Does adhering to the network perspective mean the end of factor analysis and other techniques associated with the more traditional perspectives on personality? No. Within the network perspective, factor analysis may become a useful technique for identifying groups of closely connected components. In fact, in special cases, it may be possible to estimate certain network parameters through factor analysis because groups of reciprocally connected components can behave exactly as predicted under a factor model (Van der Maas et al., 2006). As such, we do ‘not’ object to latent variable ‘modelling’ in which conditional independencies implied by a statistical model are investigated and tested. Also, we readily acknowledge that some of the hypotheses that follow from the network perspective could in principle be tested with latent variable techniques (e.g. testing the influence of genes on individual personality items with independent pathway models) nor do we deny that if some relatively unexplored areas of the latent variable realm would be more extensively cultivated in personality research (e.g. intra-individual factor modelling over time and state–trait modelling within a latent variable framework; Steyer, Schmitt & Eid, 1999), the latent variable model might be equally capable of accommodating certain phenomena compared with the network perspective (e.g. accommodating both inter-individual differences and day-to-day variation).
The question of which techniques are capable of doing what is, in our opinion, not the one that should matter most in personality research. There is and should be no arms race at the level of the (future) technical accomplishments of both models. What matters most is which perspective provides the most plausible account of how personality arises: do traits cause cognitive, affective and behavioural components or do traits emerge from complex interactions between these components? How can future research help in finding an answer to this pivotal question? Given the current lack of methodological sophistication of the network models, the most likely frontrunner in terms of empirical research will be time-series analysis of intra-individual data. Such data can, for example, be collected by assessing individuals' current thoughts, feelings and behaviours at many consecutive time points (for example by means of an experienced sampling protocol, which has been developed in considerable detail in clinical psychology). If time-series analysis of such data would show that, within individuals, personality components have a (bi)directional influence on one another, then this would be strong evidence in favour of the network hypothesis and against the latent variable hypothesis. Another research strategy might be an inter-individual approach, in which one would experimentally test whether manipulating one personality component has an effect on another personality component.
In our view, the reification of factors such as extraversion as causes of individual behaviour is unnecessary and unwarranted in the case of personality. That is, we do not object to latent variable ‘techniques’ but we do object to a latent variable ‘theory’ in which the measurement model with a common cause structure is interpreted as evidence for latent causal entities that operate in the minds of individuals, causing all sorts of cognitive, affective and behavioural patterns (see also Borsboom, 2008b). Human behaviour simply does not appear to work this way: it is not extraversion that causes party going; extraversion emerges out of liking parties, liking people and enjoying conversation.
The work of Angélique Cramer and Denny Borsboom was supported by NWO innovational research grant no. 451-03-068. The work of Sophie van der Sluis was supported by NWO/MaGW VENI-451-08-025 and VIDI-016-065-318. The data used for the genetic analyses described in this manuscript were obtained from the GAIN Database found at http://view.ncbi.nlm.nih.gov/dbgap, controlled through dbGAP accession number phs000020.v2.p1. Samples and associated phenotype data were provided by the Netherlands Study of Depression and Anxiety (NESDA) and the Netherlands Twin Register (NTR).
We acknowledge that personality inventories tend to measure self-concept, a person's view of one's own personality that might, to some extent, deviate from one's actual, objective personality. Because the network perspective is undecided concerning whether or not personality networks should be solely based on objective personality, or on both objective personality and self-concept, using personality items as a starting point for defining personality components is sensible, also given the lack of viable alternatives. Future experimental research with a focus on elucidating whether thoughts/feelings/acts, and their mental representations, have unique effects on other personality components might prove beneficial in refining personality networks in terms of what components they should contain: objective ones or also their mental representations. However, we note that current latent trait models often do equate self-concept and objective personality: the personality literature shows an abundance of statements concerning traits (e.g. women are more neurotic compared with men, suggesting an objective difference); the evidence for which is often based on personality inventory items (a more appropriate statement would then be as follows: women's self-concept of their personalities tends to include more neurotic features compared with men).
Please note that for this graph, and the other networks that are presented in this paper, the positions of the nodes in the graph are not identified. That is, by using a force-embedded algorithm, the graphs are a two-dimensional representation of networks that are multi-dimensional. In this representation, the position of a node is defined relative to other nodes in the network. The resulting distance in two dimensions between two nodes does not represent the correlation but, rather, is an approximation of the distances in the multi-dimensional network.
For the sake of simplicity, this papers focuses on activation as a dichotomous characteristic of a personality component: it is either “on” or “off”. However, it is also possible to define activation as an ordinal or continuous characteristic, in which components’ activation varies along a scale. In this way, components in a personality network can be active with a certain ‘intensity’.
Some may argue that height is not an appropriate example of a latent variable. However, see these papers for an explication of the point that latent variables are variables that are not directly observed (i.e. we cannot observe directly whether someone is 5.1 ft or 5.2 ft). We therefore need measurement instruments to quantify such variables. Also, for example, Harman (1960) argued that a latent variable is an underlying variable that helps explain why certain other variables correlate (i.e. two methods to measure height in Bob correlate because they are caused by the same underlying variable, namely Bob's height). In both non-formal views of the theoretical status of a latent variable, height is an appropriate example.
In every unidimensional latent variable model, the sum score has a monotonic likelihood ratio with the latent variable, thereby rendering the sum score a better approximation of the latent variable than a single item.