Much exploratory network research has been conducted in the field of psychopathology network research: but where do we go from here? In this section, we will discuss some future perspectives, structured into clinical and methodological research.
Clinical research
From a clinical perspective, we suggest to investigate four topics. First, the network framework generates specific hypotheses about treating disorders that should be explored. In treating comorbid disorders, such as MDD and GAD, for example, targeting bridge symptoms that transfer influence from one part of the network to the other should be the strategy of choice. A related hypothesis is that targeting central symptoms should reduce patients’ symptomatology [
46].
3 As the majority of research on finding possible targets for intervention is based on cross-sectional data, it is unclear whether an undirected edge between symptoms A and B implies A → B, A ← B, or A ⟷ B.
4 Longitudinal analyses allow for an estimation of directed networks which reveal the direction of the association between symptoms, such as A and B (e.g., [
51]), and present a more promising route to investigate possible targets for clinical intervention.
Second, and related to the previous point, it should be investigated whether intervening on central symptoms will actually bring benefits to patients. Although studies have collected ESM data in therapeutic settings to provide feedback on patterns of affect [
58‐
60], these data have not been analyzed using network models to derive, for instance, the most central symptoms—and a large crowdsourcing study that does provide feedback via personalized dynamic networks does so only outside a therapeutic setting [
61]. Merging these two approaches may provide valuable insights, and we are aware of one such pioneering case study that investigated personalized feedback based on network models within a therapeutic setting [
62]. In addition to treatment as usual, the patient received feedback on symptom dynamics and explored the feasibility, acceptability, and usability of such an integrated individualized network approach. This initiated a therapeutic dialogue about possible causes of treatment resistance and may provide new directions towards personalized medicine [
63,
64]. While it may not always be feasible or possible to target a specific symptom, establishing that the network framework provides good explanatory and predictive models of psychopathology may imply the need for developing new approaches for targeting specific symptoms.
Third, it would be worthwhile to apply the network perspective to yet unexplored mental disorders. The temporal dynamics of symptoms of Binge Eating Disorder (BED), for example, may be a suitable candidate [
12]. One causal pathway could be between the symptoms eating until one feels uncomfortably full and feeling disgusted, depressed, or guilty, which could provide insights into risk factors for the development of BED episodes.
Fourth, besides looking at the interactions of problems (e.g., symptoms), studying factors that contribute to resilience may be worthwhile pursuing [
65,
66]. Investigating the role of protective factors in psychopathology networks might inform us how these two opposing forces relate to each other, and eventually inform clinical practice. For example, Alice may benefit from more social interactions in case social isolation leads to sad mood, while Bob may benefit from physical activity in case sad mood is preceded by lack of activity.
Methodological research
Exploring the above questions relies on the accurate and reliable
estimation of psychopathological networks. When patients apply for treatment,
5 there is often a waiting period in which one could assess the emotion and symptom dynamics with modern phone technology within an idiographic momentary assessment study, and results could inform treatment. Similarly, relapse prevention in remitted patients may benefit from repeated assessment of core symptoms and related factors over time to foresee relapse in an early phase and take preventive measures to counteract its course. This all sounds promising, but before this can be put into effect, there are some methodological issues that need to be addressed of which we will discuss three.
A first issue is what variables to study in psychopathological networks. While cross-sectional network studies have focused on analyzing associations among symptoms, ESM studies have focused on mood states, such as sadness, happiness, anxiety, or anger [
4,
38,
49,
67]. It is unclear at present what level of variables is best to study psychopathology.
A second issue is the time frame on which to measure symptoms or emotions. In most ESM studies, the time frame between measurements is a few hours. However, do symptoms or affects change within hours or minutes or days? This might differ for different pairs of symptoms: experiencing somatic arousal (e.g., increased heart rate and sweating) might lead to anticipating a panic attack [
43], which will occur within minutes. Sleep problems, on the other hand, might build up for a few days before influencing a person’s irritability. It is currently unknown what the best timeframe is to capture dynamics.
Third, an important point is the generalization of group-level results to the individual level, since many group-level network studies have implied that the identified network structure of the population is more or less reflective of the networks of all individual participants (e.g., [
68,
69]). A well-known example of this phenomenon, known as Simpson’s Paradox, is the speed–accuracy tradeoff. At a group-level, a negative relationship exists between typing speed and typing accuracy: people with higher typing speed make fewer errors, likely because experience leads to faster typing and fewer mistakes. At the individual level, however, a person who types faster will make more, not less errors [
70]. While this is an extreme example—it seems unlikely that symptoms of mental disorders are predominantly positively associated at group-level, but negatively in the individual—we currently do not know to what extent group-level networks differ from individual networks [
43]. A related point was made by Bos and Jonge [
71] and Bos and Wanders [
42] who warn that between-person effects should not be confused with within-person effects. Taken together, this implies that we need future studies that investigate to which degree idiographic networks match group-level networks, and to disentangle between-person from within-person effects.
Finally, numerous network papers analyzed data that contained a skip structure. This is often the case when large populations are screened via the DSM diagnostic criteria. For a diagnosis of MDD, for instance, subjects need to endorse at least one of the two core symptoms depressed mood or anhedonia. If that is not the case, the remaining seven MDD symptoms are skipped. In statistical analyses, such skipped items are usually recoded as 0s (e.g., [
10,
19,
53]), but just because someone does not endorse the core symptoms does not mean that the person cannot exhibit other MDD symptoms. The recoding of missing data to 0s may pose a considerable problem, because it introduces spurious correlations among items (for many people, the seven remaining items will be coded as 0s and thus be highly correlated, although this may not reflect the true correlations among items). Although Boschloo et al. [
18] showed similarity of the network structure based on the original data with 49% missing and a subsample with less than 20% missing, it still may have introduced bias. Future research is required to investigate imputation strategies for skip data that go beyond recoding them as 0s.