Personalized Models of Psychopathology as Contextualized Dynamic Processes: An Example from Individuals with Borderline Personality Disorder

Abstract

Objective:

Psychopathology research has relied on discrete diagnoses, which neglects the unique manifestations of each individual’s pathology. Borderline personality disorder combines interpersonal, affective, and behavioral regulation impairments making it particularly ill-suited to a “one size fits all” diagnosis. Clinical assessment and case formulation involve understanding and developing a personalized model for each patient’s contextualized dynamic processes, and research would benefit from a similar focus on the individual.

Method:

We use group iterative multiple model estimation, which estimates a model for each individual and identifies general or shared features across individuals, in both a mixed-diagnosis sample (N=78) and a subsample with a single diagnosis (n=24).

Results:

We found that individuals vary widely in their dynamic processes in affective and interpersonal domains both within and across diagnoses. However, there was some evidence that dynamic patterns relate to transdiagnostic baseline measures. We conclude with descriptions of two person-specific models as an example of the heterogeneity of dynamic processes.

Conclusions:

The idiographic models presented here join a growing literature showing that the individuals differ dramatically in the total patterning of these processes, even as key processes are shared across individuals. We argue that these processes are best estimated in the context of person-specific models, and that so doing may advance our understanding of the contextualized dynamic processes that could identify maintenance mechanisms and treatment targets.

Contemporary psychiatric diagnostic systems, such as the Diagnostic and Statistical Manual of Mental Disorders (DSM, American Psychiatric Association, 2013), have had an outsized impact on clinical research. Although a shared nomenclature and consensual definition of psychopathology is advantageous, misalignment between a nosology’s descriptions and clinical phenomenology risks constraining new insights. The DSM’s psychiatric taxonomy uses discrete categorical syndromes, which are polythetic (i.e., require some subset of a list of possible symptoms for diagnosis) in an effort to accommodate the heterogeneity in clinical manifestations. As a result, two individuals can share a diagnosis of borderline personality disorder (BPD), for example, despite having only one symptom in common. Accordingly, although traditional diagnostic constructs may hold some heuristic clinical value, they are insufficiently precise for mechanistic clinical research. We emphasize the limitations for clinical research, because clinicians already look beyond diagnoses to consider the specific presenting problems of the individual in the consulting room (e.g., Hamilton et al., 2008; Kuyken, Padesky, & Dudley, 2011). However, to the extent that research and clinical practice are also misaligned, translation will be stymied. We argue that this problematic feature of categorical diagnosis can be resolved by shifting from the diagnosis to the individual as the unit of analysis in psychopathology research. This requires viewing psychopathology as ensembles of contextualized dynamic processes that involves multiple mutually influencing systems (i.e., affective, interpersonal, behavioral, cognitive, biological).¹

The current study adopts this approach in examining interpersonal and affective processes that are centrally relevant to the diagnosis of BPD. BPD is an ideal exemplar for illustrating this approach to studying transdiagnostic processes in psychopathology because (a) the diagnosis emerges from the confluence of dysregulation in affect, interpersonal relating, and behavioral control, (b) and many of the symptoms articulate contextualized dynamic processes. We use recently developed statistical methodology that builds individualized models of within-person processes and then identifies shared elements, allowing for heterogeneity in pathology while still locating commonalities between individuals. This is consistent with a broader push towards personalized diagnosis and intervention in medicine, (Collins & Varmus, 2015), prevention science (Ridenour, 2019), education (Reber, Canning, & Harackiewicz, 2018), and psychology (Fisher, 2015; Molenaar, 2004). We provide an empirical illustration of the importance of focusing on the transdiagnostic elements of heterogeneity, variation across time and circumstance, and multiple reciprocal domains of functioning in understanding BPD and psychopathology more broadly.

Traditional psychopathology research has prioritized nomothetic designs, where individuals are grouped within diagnostic categories (or, at a minimum, gives them a dimensional score derived from the polythetic criteria), without little regard to how similar symptoms are across individuals (Barlow & Nock, 2009; Fisher, 2015). The idea behind the use of nomothetic research practices is that identifying differences in groups (e.g., those in the sample with BPD and those without) allows for generalization to the wider population (e.g., those in the population with BPD and those without). However, generalization based upon groups such as those becomes problematic to the extent that individuals are not homogenous in their pathological processes. For example, it is problematic to make inferences about all individuals with a BPD diagnosis based on a sample in which, for example, some individuals have primary symptoms of affect dysregulation while others have primary symptoms of interpersonal dysfunction. Aggregating data from individuals with diverse presentations is imprecise and likely a major contributor to slow progress in the identification of etiological and maintenance mechanisms as well as maximally effective treatments. Although the diagnosis of BPD conveys some important clinical information, the clinical reality is that those who share a diagnosis that differs in its expression require further assessment to identify the important targets for intervention.

An alternative strategy is to adopt a person-specific approach to clinical research that treats the individual patient as the unit of analysis (Barlow & Nock, 2009; Molenaar, 2004; Fisher, 2015; van Os et al., 2013; Wright, 2011; Wright & Woods, 2020). Focusing on the person, rather than on the diagnostic group, has many advantages, such as being a closer match for clinical theories about psychopathology that outline how to evaluate the specific processes occurring within a given patient and then provide treatment recommendations based upon those processes (e.g., Clarkin, Yeoman, & Kernberg, 2005; Linehan, 1993, Pincus, 2005). Of course, given the established diagnostic zeitgeist, these are often framed within the nomothetic understanding of the diagnostic group. Thus, we are not arguing for the abandonment of nomothetic research—but we do recommend a rebalancing that focuses on the individual as the unit of analysis in addition to traditional nomothetic research (Roche, Pincus, Rebar, Conroy, & Ram, 2014; Wright et al., 2015). Nevertheless, using the person as the unit of analysis for psychopathology research may facilitate rapprochement in the persistent research practice divide (Stricker & Goldfried, 2019). This is because personalized models of psychopathology approximate the clinicians understanding of patients as distinct individuals rather than as the “average” patient from a nomothetic sample (Fisher, 2015). This may be especially relevant for complex syndromes such as BPD, which emerge at the confluence of multiple domains of dysfunction, as compared to more narrowly defined diagnoses such as simple phobia or panic disorder that are organized around specific processes with well-established treatments. Because the ultimate goal of psychopathology research is the amelioration of human psychological suffering, research needs to be useful to the clinician, not just other researchers. It is a luxury available only to researchers to examine individuals with minimal comorbidities, whereas the frontline clinician must treat each patient’s unique presentation in all of its complexity.

Nomothetic research is founded on the differences (i.e., variability) between individuals, whereas idiographic research requires studying the variability within an individual across time and circumstance (Allport, 1937; Nesselroade & Molenaar, 2016). By taking the variability in states across time, and sampling important variables in each state, a personalized model can be estimated based on the individuals own processes (Fisher, 2015). This approach views individuals as complex systems and psychopathology as ensembles of maladaptive contextualized dynamic processes (e.g., van Os et al., 2013; Wichers, Wigman, & Myin-Germeys, 2015). This is precisely how the DSM defines BPD as shown in the core features of the disorder: frantic efforts when abandonment is perceived, unstable interpersonal relationships, unstable sense of self, affective instability, and so forth. This is consistent with a growing body of empirical research that has sought to systematically measure these processes using intensive longitudinal research designs in BPD (e.g., Ebner-Primer et al., 2007; Hepp et al., 2017; Miskiewicz et al., 2015; Sadikaj et al., 2013; Scott et al., 2017; Trull et al., 2008) and personality disorders more broadly (Wright & Simms, 2016). Notably this process-based research has revealed that even cardinal features of BPD, such as instability in affect does, in fact, transcend traditional diagnostic boundaries (Santangelo et al., 2014). Yet, here too, dynamic processes are often compared across diagnostic and control groups rather than examining the potentially rich heterogeneity in these processes within a given group, with some few exceptions we will discuss below (Ellison et al., in press; Fisher, 2015; Fisher et al., 2017; Lane et al., 2019; Wright et al., 2019).

Returning to BPD, it is likely that any two individuals with a BPD diagnosis will not share identical pathological processes (Wright et al., 2016). This may seem an obvious point but the implications for treatment and understanding of pathology are important. Consider two hypothetical individuals with BPD diagnoses—one with primary affect regulation problems (Patient A) and one with primary interpersonal difficulties (Patient B). In the person-specific research paradigm, it is assumed that all of the functional domains are mutually interrelated and influence each other to some degree or another. It could be the case, as would be assumed in a nomothetic treatment study framework, that these individuals share the same underlying system of functional associations across domains. However, once we have shifted to the person-specific paradigm, we are agnostic about the assumption that these individuals do, in fact, share the same pattern of connections among domains. For example, Patient A’s affective system may exert a much stronger influence on all other domains than Patient B’s, which makes dysfunction in this domain even more impairing than that for Patient B, whose affective domain is not as central to the overall system even if impaired relative to an individual without personality pathology. Instead of assuming homogeneity among processes and comparing across groups, we believe it would be fruitful to treat individuals as unique ensembles of processes and evaluate whether and when homogeneity exists rather than assuming similarity in processes a priori.

Here, as others have done as well (Fisher, 2015), we advocate for formal personalized quantitative models of each person’s psychopathology. Although our emphasis is on applications in psychopathology research, this general approach offers direct translation to practice via precision assessment (Roche et al., 2014; van Os et al., 2013). Some readers may question whether personalized models are needed, or whether the heterogeneity in processes could be adequately accommodated via multilevel modeling (i.e., mixed effects models) with random effects. Multilevel models have, at times, been cast as a blend of nomothetic and idiographic approaches (Conner et al., 2009). This is not strictly correct, though, and is likely insufficient to accommodate the system-level heterogeneity of an individual’s psychopathology. By estimating random effects, multilevel models do accommodate heterogeneity in a given association across individuals in the sample. At the same time, these random effects assume normal distributions and when these are violated (e.g., with extreme outliers as might be the case in psychopathology) severe bias can imperil accurate inference. Furthermore, multilevel models do not simultaneously estimate a full network of associations among all included variables for each person, and therefore do not provide a full accounting of an individual’s “system” of psychopathology. As such, fully personalized models are likely needed to best approximate each individual’s patterning and structure.

Other readers may be skeptical about the ability of personalized models based on idiographic research to inform population-level models of psychopathology (Beltz et al., 2016). After all, traditional n-of-1 studies do not typically allow for easy comparison between cases, and generalization is not typically an explicit goal of most idiographic research. These are valid concerns—generating nomothetic conclusions from idiographic research requires standardization in testing and statistical methodology for analyzing individualized datasets (Wright & Zimmermann, in press). However, one method capable of identifying homogeneous processes shared by person-specific models is group iterative multiple model estimation (GIMME; Gates & Molenaar, 2012; Lane et al., 2019).

GIMME is currently based on unified structural equation modeling (Kim et al., 2007), which integrates vector autoregression and structural equation modeling.² Unified structural equation models estimate both the contemporaneous and lagged directed effects (i.e., regression paths) among a set of time-series variables. Unlike multilevel models (i.e., mixed effects models), which estimate a set of average or fixed effects and random variability around those effects, GIMME estimates a person-specific model for each individual (see Gates, Lane, Varangis, Giovanellow, & Guskiewicz, 2017, for thorough treatment of GIMME’s estimation methods). However, GIMME also searches for paths that that are present in the majority of a sample’s models. Those paths are then incorporated in each individual model in successive iterations of estimation of each individual’s model. By searching for shared paths, and giving those priority in each model’s estimation, GIMME bridges the idiographic and nomothetic levels of analysis.

GIMME’s estimation proceeds in several stages (Gates & Molenaar, 2012; Lane et al., 2019). First, it examines all possible lagged and contemporaneous paths and identifies those that would improve the model fit for the majority of individuals in the sample (the exact percentage is modifiable by the investigator). It is important to highlight that, at this first stage (i.e., the group or whole sample level), GIMME is not aggregating across all individuals, but rather treating each participant as a unique sample; paths identified as significant at the group level improve the individual model fits for the majority of the person-specific models. In this way GIMME is robust to the influence of outlier individuals. Paths added to the group level structure must statistically improve the individual model fits for the majority of the person-specific models and may vary for each individual in terms of the strength and sign of the relationship between any two variables.

Using the pre-selected group paths as a foundation (i.e., they are freely estimated in each model), GIMME then performs individual level searches for significant effects among the remaining potential paths. This process is repeated for each participant in the sample until each individual-level model achieves excellent fit. To prevent overfitting, the search process is stopped once the individual’s model shows excellent fit (Gates & Molenaar, 2012) and, in addition, the process contains a Bonferroni correction to protect against false positives. A number of simulations suggest that the algorithm’s emphasis on parsimony consistently avoids false positives (Gates et al., 2011; Gates & Molenaar, 2012; Lane et al., 2019). At the end of the estimation procedure, each individual in the sample has unique estimates for the group paths as well as their own personal paths. By using this estimation procedure, GIMME is capable of overcoming the obstacles outlined above: it can reliably estimate individualized models while identifying homogenous processes, incorporate the dynamic aspects of psychopathology, and incorporate multiple psychological domains and their mutual influence on each other.

The current study joins a fast-growing body of research that has emerged in recent years using a person-specific or idiographic approach to understanding psychopathology (see Wright & Woods, 2020 for a review). Although the idiographic or n=1 approach to psychopathology has a long history (Barlow & Nock, 2009; Haynes et al., 2009), only recently have developments in data capture and modeling facilitated large-scale idiographic studies that go beyond the evaluation of a few individuals’ models at one time. However, despite these advances, studies involving more than a handful of participants remain rare. This is especially true when considering studies that seek to model psychopathology as a multidimensional system of interacting domains of functioning. What few studies exist have either focused on internalizing disorders, or, as we do here, personality pathology. For instance, in a series of studies, Fisher and colleagues (Fisher, 2015; Fisher & Boswell, 2016; Fisher, Newman, & Molenaar, 2011; Fisher et al., 2017) have estimated idiographic models of the associations among symptoms of generalized anxiety disorder and depression sampled multiple times per day. Most relevant for the current context, using a sample of individuals (N=40) diagnosed with either major depressive disorder or generalized anxiety disorder, Fisher and colleagues (2017) estimated idiographic models very similar to the uSEM approach that GIMME uses on momentary assessments of symptoms depression and anxiety. The major conclusion was that despite the focus on two highly overlapping diagnoses, the individual models were decidedly heterogeneous, suggestive of unique dynamic processes that differed markedly across participants. To bridge back to nomothetic conclusions, Fisher and colleagues averaged the various parameters and presented sample average networks.

In a more recent study, Bosley and Fisher (2019) examined the network of associations among worry, positive affect, and affective dampening in a sample of undergraduates with elevated anxiety (N=96). Again, the key takeaway was that these processes are highly heterogeneous in nature across individuals. In an effort to compare the individual results to a nomothetic model, all participants’ data were aggregated and contributed to a single model. A limitation of this aggregation approach is that it confounds between and within-person variability in a single model making it very difficult to directly compare to the idiographic models that are based exclusively on within-person variation (see e.g., Borkenau & Ostendorf, 1998 for a similar approach). This serves as a useful comparison to the GIMME approach, which estimates exclusively within-person effects thus retaining interpretive clarity, and for which group effects are those that are significant in a majority of the sample. Thus, a single or handful of large outlying effects could not “create” a shared effect if it was uncommon. More broadly, this approach does not conflate within- and between-person variance in what has sometimes been termed a “compromise” or “smooshed” effect.

There have also recently been several idiographic studies of personality pathology, and all have used the GIMME approach. Two studies have used daily diary assessments in a sample (N=90+) of individuals diagnosed with any personality disorder, with one using daily interpersonal behavior, affect, stress, and functioning (Wright et al., 2019) and the other using daily measures of the transdiagnostic psychopathology domains of negative affectivity, detachment, disinhibition, and hostility (Dotterer et al., 2019). Similar to Fisher et al. (2017), in each sample heterogeneity was the rule, and only one group-level path (i.e., shared by the majority of participants) was found in one of the studies (stress predicting negative affect on the same day). Dotterer and colleagues (2019) also showed that some of the idiographically estimated paths incremented average daily levels in predicting traditional diagnosis severity. Unlike the internalizing work reviewed above, in each of these studies, inclusion was predicated on having any personality disorder diagnosis, such that the heterogeneity in processes was expected and, in some respects, unremarkable.

To provide a more focal test of heterogeneity in processes in personality pathology, Lane et al. (2019) conducted the first ambulatory assessment-based GIMME analysis of individuals (N=35) with BPD diagnoses selected from the same daily diary sample. The analysis included affective (depressed, anxious, angry) and disinhibition (impulsivity, urgency) related variables, as well as a measure of emptiness assessed at the daily level. Despite the narrower focus on a shared diagnosis, the results again highlighted that the overall structure of processes (i.e., associations among variables within and between days) was highly heterogeneous. This was determined by the fact that there was only a single group-level path and each possible path was present in only a minority of models and varied in strength. In a follow-up study, Ellison and colleagues (in press) used GIMME to examine associations among affect, impulsivity, and identity disturbance items in a mixed sample of BPD and anxiety disorder patients. Their findings were highly consistent with the earlier results of Lane et al. (2019) despite differences in sampling rate (momentary vs. daily), included variables, and participant diagnoses.

The current study builds on the existing research on personalized models of psychopathology in several ways. First, to illustrate the transdiagnostic value of the personalized approach, we estimate models in a mixed clinical and community sample as well as a subsample of individuals diagnosed with BPD to provide a direct comparison of the heterogeneity within a specific diagnosis and across diagnoses. Nevertheless, consistent with Lane et al. (2019) and Ellison et al. (in press), we anticipate that the BPD sample will exhibit highly heterogeneous models across individuals as evidenced by few group-level paths. BPD is ideal for showcasing the benefits of person-specific approaches because of its complexity as a pathology with dynamic, multiple relationships between social, affective, and cognitive domains. Indeed, all major contemporary theories of BPD (e.g, Clarkin et al., 2006; Fonagy & Luyten (2009); Linehan, 1993) converge on this idea, while diverging as to the relative contributions of each domain to the pathology’s manifestation (Gunderson et al., 2018). As noted above, the BPD patients who present for treatment vary widely in their presentation with the ascendant problem domain varying from patient to patient. The current study features data from the interpersonal and affect domains. Using GIMME allowed us to demonstrate how these domains influence each other in heterogeneous ways and to identify both commonalities and differences among those with the same putative diagnosis. Identical analyses in the larger, mixed sample also underscore the heterogeneity of cross-domain processes—favoring person-specific models—and allow us to demonstrate how interindividual differences in paths between different affects and interpersonal perceptions can be linked to pre-existing cross-sectional measures that matter to clinicians.

Second, in addition to the novel use of a BPD sample for person-specific analysis, we also extend previous work by using event-focused (i.e., social interaction) data rather than fixed or random-interval sampled data, which has most often used in idiographic research (e.g., Fisher, et al., 2017). This allows us to focus in on the context (i.e., situation) in which much of the maladaptively and impairment is anticipated to manifest. As a result, this enabled us to sample and incorporate important social processes, including perceptions of others’ behavior (i.e., perception of situational features) and one’s own interpersonal behavior and affect in the moment, that are broadly relevant across diagnoses (Pincus & Wright, 2011).

Third, we used the gimme R package’s (Lane, Gates, Fisher, & Molenaar, 2019) new exogenous variable feature to adjust for the effects of time (i.e., detrend) in our path estimates. This new feature is valuable because it reduces the need for extra and oftentimes cumbersome data processing to remove unwanted cyclical and other time effects from the data prior to analyses. In addition to raw time, we also included lag distance as a way to accommodate differences in spacing between reported social interactions. It is possible that the time between (i.e., lag) interactions may be related to processes of avoidance or withdrawal for some individuals. Finally, we illustrate one strategy to link the personalized model features estimated by GIMME to common measures of transdiagnostic processes relevant to BPD, such as emotion regulation, interpersonal problems, impulsivity, and dyadic adjustment using a data driven approach that can accommodate large numbers of predictors and outcomes. Specifically, we predict scores on traditional dispositional measures from the GIMME-estimated paths using least absolute shrinkage and selection operator (LASSO) regression.

Method

The current study was approved by the University of Pittsburgh Institutional Review Board (PRO12030125).

Participants

Participants were recruited as a part of a larger study using flyers in local psychiatric clinics. The parent study (N = 207) was designed to investigate the effects of borderline personality disorder (BPD) on functioning in romantic couples (Beeney et al., 2019; Lazarus et al., 2018). Individuals with past or current bipolar disorder or psychosis were excluded from the study. The current study defined individuals with BPD as those who met diagnostic threshold for the disorder on the Structured Interview for DSM-IV Personality (SIDP-IV; Pfhol, Blum, & Zimmerman, 1997). In line with recommendations from simulation studies of GIMME’s ability to recover group and individual paths (Lane et al., 2019), we limited our analyses to those with at least 60 responses (n = 84). We further excluded participants with no variance on at least one of the variables of interest because variability is needed to calculate associations at the person level (n = 6). Among these participants (n = 78), the total number of reported interactions was 6,972, with a median of 85 (Range = 60 – 161). Participants were predominantly female (61.5%) and white (52.6%). The remainder identified as African American (28.2%), mixed race (12.8%), and Asian (6.4%). The average age was 31.4 (SD = 7.3).

In the subset of this sample that were assigned to the BPD subgroup (n = 24), most identified as female (66.7%). Race was mixed: 37.5% identified as White, 29.2% as Black, 25.0% as mixed race, and 8.3% as Asian. The mean age of this group was 30.1 (SD = 6.9).

Procedure

Participants completed baseline assessments prior to undertaking a 21-day EMA protocol. The EMA data was collected using smartphones provided to participants which had a pre-installed custom application designed for the study. Participants were asked to complete a questionnaire from the application immediately following each interpersonal interaction lasting at least 10 minutes. The questionnaire assessed aspects of the interpersonal interactions as well as the participants’ affect and behavior. Participants were compensated up to $205 for their cooperation with the study procedures; $40 for completing the structured interviews and up to $165 for daily submissions of questionnaires.

Measures

Personality Pathology –

Participants were assessed for personality disorder symptoms using the SIDP-IV (Pfohl, Blum, & Zimmerman, 1997). Interviewers were trained clinicians with a master’s or doctoral degree. Each PD criterion was rated using a 0–3 scale. We used the symptom present/absent scoring method to determine BPD diagnostic threshold and inclusion in subsequent subsample analyses. Seven clinical evaluators independently scored the SIDP-IV interview from videotape for 5 participants to establish interrater agreement, which was high for BPD symptoms (intraclass correlation = .98).

Momentary Interpersonal Behavior –

Participants assessed their own interpersonal behavior using the Social Behavior Inventory (SBI; Moskowitz, 1994). The SBI is a checklist of 46 behaviors designed to capture the interpersonal dimensions of dominance and affiliation (i.e., warmth). Participants indicated whether each behavior was emitted during the interaction using yes/no response choices. In the SBI, each pole of each interpersonal dimension is assessed separately. Dominance is assessed in terms of Dominant behaviors (e.g., “took the lead; expressed an opinion”) and Submissive behaviors (e.g., “gave in; did not say how he/she felt”). Warmth is assessed via Agreeable behaviors (e.g., “smiled/laughed; pointed out agreement”) and Quarrelsome behaviors (e.g., “confronted others; showed impatience”). Participants were presented with 12 of 46 possible items each time they completed an interpersonal questionnaire. Following previous work (Sadikaj et al., 2013), four forms of SBI items were developed, each with three items for each pole of each interpersonal dimension. Each interpersonal dimension was scored by subtracting the sum of the negative pole items from the sum of the positive pole items (i.e., Dominance = Dominant − Submissive; Warmth = Agreeable − Quarrelsome).

Participants rated their interaction partner’s interpersonal behavior using seven items adapted from the SBI. Two items assessed each pole of the two dimensions, with the item “s/he criticized” being shared by both the dominance pole and the coldness pole. As with the ratings of self-behavior, dominance and warmth scores were calculated by subtracting the sum of the negative pole items from the sum of the positive pole items.

Affect –

Momentary affect was assessed using 27 items from the Positive and Negative Affect Schedule-Extended Version (PANAS-X; Watson & Clark, 1999). Participants rated the extent to which they experienced each affect on a 1–5 scale anchored from “very slightly or not at all” to “extremely.” Subscale scores for fear (6 items), hostility (6 items), sadness (5 items), and positive affect (10 items) were calculated as the means of each subscale’s respective items.

Difficulties Regulating Emotion –

Dispositional emotion dysregulation was assessed using the Difficulties with Emotion Regulation Scale (Gratz & Roemer, 2004). This 36-item measure features 6 subscales: Nonacceptance of Emotional Responses, Difficulty engaging in Goal-Directed Behavior, Impulse Control Difficulties, Lack of Emotional Awareness, Limited Access to Emotional Regulation Strategies, and Lack of emotional Clarity. Items were rated on a 5-point scale ranging from 0 to 4.

Impulsivity –

Impulsive behavior was measured using the UPPS Impulsive Behavior Scale (Lynam, Smith, Whiteside, & Cyders, 2006). This 59-item measure assesses five factors of impulsivity. Participants respond to items using a scale ranging from 1, “agree strongly,” to 4, “disagree strongly.” For the purposes of the current study, mean scores for the five subscales were calculated.

Interpersonal Problems –

Chronic difficulties with interpersonal interactions were assessed using 64 items from the Inventory of Interpersonal Problems (Alden, Wiggins, & Pincus, 1990). Responses are on a 5-point scale ranging from 0, “not at all,” to 4 “extremely distressing.” Circumplex scoring was applied to the items and the two axes of problems with dominance and affiliation were calculated as well as a global general distress dimension.

Dyadic Adjustment –

Quality of romantic attachment was measured using the Dyadic Adjustment Scale (Spanier, 1976). This 32-item measure assesses the extent to which romantic partners are adjusted into their relationship. Items had varying response scales that were used to score the four scales: Dyadic Consensus, Dyadic Satisfaction, Dyadic Cohesion, and Affective Expression.

Analytic Strategy

Personalized and group-level models of situational affect (fear, hostility, sadness, and positive affect) and interpersonal behavior (perceptions of other dominance and affiliation, self-rated dominance, and self-rated affiliation) were generated using the R package gimme (Lane, Gates, Fisher, & Molenaar, 2019). In addition, we used gimme’s new exogenous variable capabilities (Arizmendi, Gates, Fredrickson, & Wright, 2019) to include the linear effect of time (to detrend any variable that exhibited significant increases or decreases over the course of the study) and time lag between reported interactions. The variables of time and lag could not be predicted by other variables, and could only predict other variables. Group-level paths were defined as those that significantly improved model fit in 51% of the personalized models. As estimated in this study, the uSEMs assume multivariate normality, the associations among the observed variables are not caused by latent variables, and that a lag of one observation can accommodate any cross-temporal effects. Violations of these assumptions can lead to bias in effects. Alternative specifications of GIMME can modify or relax these assumptions (e.g., Dotterer et al., in press; Gates, Fisher, & Bollen, in press). Because these analyses were conducted on data that were not specifically designed for personalized models, only a minority of participants (n=78) from the parent sample (N=207) had sufficient observations to be included (i.e., 60 or more observations, sufficient variance in each variable). Therefore, our sample is not necessarily representative of the broader sample, and want to emphasize that we view these analyses as exploratory and in the service of hypothesis generating, rather than confirmatory. GIMME was first applied in the subsample of participants with a diagnosis of BPD. We next expanded the analyses to include all participants who met the inclusionary cutoffs described above to examine whether paths evident in individuals with BPD were evident in the full sample (i.e., transcended diagnostic boundaries).

After obtaining each individuals’ model in gimme, LASSO regression analyses were performed using the GLMNET package in R (Friedman, Hastie, & Tibshirani, 2010) to probe and complement within-person results obtained from gimme. LASSO is a regression method for arriving at a sparse solution through use of a penalty. With the penalty, many regression coefficients are shrunk to zero, suggesting that those variables are unimportant to explaining the outcome (Friedman et al., 2010). The predictor variables were the individuals’ beta estimates for each regression path in their personalized model (i.e., estimates of lagged and contemporaneous associations among variables). The dependent variables were a variety of trait-level characteristics selected to identify the major transdiagnostic domains relevant for borderline personality pathology including difficulties in emotional regulation, interpersonal problems, and impulsivity. Trait-level variables were included as outcomes in separate adaptive LASSO models. Again, we present these analyses as an exploratory only, hypothesis-generating mechanism to inform future studies designed to specifically develop personalized models of psychopathology relevant to BPD.

Finally, to compare the models of BPD and non-BPD diagnosed participants, we calculated network density degree. Here degree is defined as the number of GIMME identified associations in each individual’s model, subtracting out the autoregressive paths (which by default are included for all individuals).

Results

BPD Subsample

GIMME identified four group-level contemporaneous paths, indicating that a majority of the sample shared these associations (although note that the only criterion is that their inclusion significantly improves model fit; Figure 1; see also Table 1). The group-level model indicated that a majority of our participants were found to have significant contemporaneous associations between sadness and hostility and hostility and fear. Interpersonally, they also were found to have significant associations between perceived warmth of the other and their own warmth towards the other. Finally, a group level association between the interpersonal and affect domains was found in a contemporaneous association between hostility and perceptions of the other’s warmth.

Figure 1:

Group Iterative Multiple Model Estimation plot of shared and individual paths in the Borderline Personality Disorder subsample. Black lines indicate shared (group-level) paths; grey lines are individual paths. Number of individuals with individual path is indicated by line thickness. Solid lines represent contemporaneous paths; dotted lines represent lagged associations.

Table 1:

Number of individuals in the borderline personality disorder subsample with significant contemporaneous and lagged paths.

Contemporaneous Paths	Fear	Hostility	Sadness	PA	Dominance	Warmth	Other Dominance	Other Warmth	Laga	Timea
Fear	-	24	6	2	1	2	1	0	0	0
Hostility	6	-	24	2	1	1	2	0	1	0
Sadness	4	1	-	5	0	1	1	2	1	3
Positive Affect (PA)	4	4	3	-	0	3	0	8	1	7
Dominance	1	1	2	1	-	4	1	0	0	1
Warmth	0	3	2	2	5	-	2	24	1	0
Other Dominance	0	2	1	1	1	2	-	7	2	2
Other Warmth	3	24	0	2	1	6	0	-	0	2

Lagged Paths	Lagged Variables
	Fear	Hostility	Sadness	PA	Dominance	Warmth	Other Dominance	Other Warmth
Fear	24b	1	1	1	1	1	0	0
Hostility	2	24b	0	0	0	2	0	0
Sadness	2	2	24b	3	0	0	0	0
Positive Affect (PA)	1	2	1	24b	0	0	1	0
Dominance	1	1	1	0	24b	1	0	0
Warmth	0	1	0	1	1	24b	2	0
Other Dominance	0	1	0	0	1	0	24b	1
Other Warmth	1	3	0	0	0	1	2	24b

Note: Column variables are predictors of row variables.

^a= Lag and Time were both treated as exogenous variables and therefore do not have lagged paths in the GIMME model.

^b= Autoregressive paths are opened for all individuals by default

In addition to the shared paths, there was also significant heterogeneity across the individual models. Table 1 includes the counts of individuals with each possible path (note column variables are predictors of row variables). Out of 56 possible contemporaneous paths (excluding those related to time or lag), 46 (72%) were significant for at least one person in the subsample. Out of these 46 significant paths, a third of these (15; 32%) were significant for only one individual, demonstrating substantial heterogeneity for individuals within the same putative diagnostic category. This number rises to over half the observed paths (27; 57%) when examining paths significant for only one or two individuals emphasizing the need for a personalized approach to psychopathology.

Turning to lagged paths, as is standard in GIMME analyses, all autoregressive paths were opened for all participants by default (Lane et al., 2019). Excluding these, there were no paths shared by the majority of the sample. In fact, the lagged path which was shared by the greatest number of participants (Perceptions of other’s warmth predicting later hostility) was shared by only three participants. Out of total of 56 possible lagged paths (excluding autoregressive paths), 29 were significant for at least one person (45%), with 20 of those being limited to only one individual, again highlighting the utility of personalized models.

Mixed Diagnosis Sample

The BPD subsample showed strong heterogeneity among idiographic models within a single diagnosis; by using a larger mixed diagnosis sample we can place the BPD findings in context and (a) possibly demonstrate broader heterogeneity of general psychopathology, and (b) that these paths are transdiagnostic. In the mixed diagnosis sample, instead of five significant group-level paths, only one path was shared by the majority of the participants: contemporaneous hostility with sadness. As also might be anticipated, a larger proportion of the possible paths were significant in at least one individual’s model: 54 out of the possible 56 (96%; excluding the time variables). Out of these 54, 11 paths were significant for only one or two participants (20%), suggesting relatively lower heterogeneity of significant paths in the BPD group. Moving from contemporaneous to lagged paths, 52 of the 56 lagged paths (93%; once again excluding autoregressive paths) were significant, with 48% of those paths significant for only one or two participants (Table 2). These results are also illustrated in Figure 2, which depicts the one shared (non-autoregressive) path as a think black line, and person-specific paths as grey lines, where thickness indicates the number of individuals with each path.

Table 2.

Number of individuals in the full sample with significant contemporaneous and lagged paths.

Contemporaneous Paths	Fear	Hostility	Sadness	PA	Dominance	Warmth	Other Dominance	Other Warmth	Laga	Timea
Fear	-	25	25	9	1	3	2	4	3	6
Hostility	11	-	5	6	5	15	6	16	3	6
Sadness	13	78	-	5	2	2	2	1	5	5
Positive Affect (PA)	7	7	17	-	3	5	0	14	3	21
Dominance	2	3	3	2	-	14	5	5	2	4
Warmth	0	13	8	7	14	-	4	27	3	4
Other Dominance	1	9	3	4	2	5	-	16	4	7
Other Warmth	1	19	6	11	3	14	9	-	0	3

	Lagged Variables
Lagged Paths	Fear	Hostility	Sadness	PA	Dominance	Warmth	Other Dominance	Other Warmth
Fear	78b	5	4	1	4	3	3	0
Hostility	4	78b	7	3	3	7	3	1
Sadness	6	5	78b	3	2	1	2	2
Positive Affect (PA)	4	4	7	78b	2	0	3	3
Dominance	2	2	4	2	78b	9	2	1
Warmth	4	2	3	3	5	78b	4	1
Other Dominance	0	2	1	1	1	2	78b	1
Other Warmth	1	4	1	1	0	1	2	78b

Note: Column variables are predictors of row variables.

^a= Lag and Time were both treated as exogenous variables and therefore do not have lagged paths in the GIMME model.

^b= Autoregressive paths are opened for all individuals by default

Figure 2:

Group Iterative Multiple Model Estimation plot of shared and individual paths in the Mixed Diagnosis subsample. Black lines indicate shared (group-level) paths; grey lines are individual paths. Number of individuals with individual path is indicated by line thickness. Solid lines represent contemporaneous paths; dotted lines represent lagged associations.

Time Variables

One innovation of the current study is the use of GIMME’s new exogenous variable function. We included time since last reported social interaction (i.e., Lag) and raw time since first report as exogenous variables. Conceptually, the lag variable represents the extent to which chronological distance from the last interpersonal interaction has an effect on a given variable. There were only a small number of lag effects across the BPD (Table 1) and the full samples (Table 2). The time variable was included as a way of detrending the data within a single model, as opposed to doing so in a separate model has been done in previous work (e.g., Fisher et al., 2017). A large proportion of both the BPD subsample (30%; Table 1) and the full sample (27%; Table 2) showed significant links between raw time and positive affect.

LASSO

In an effort to bridge the idiographic models with traditional individual differences measures of transdiagnostic dimensions, we used LASSO regression to predict scales related to dysregulation of emotion, interpersonal problems, impulsivity, and dyadic adjustment from individual GIMME-estimated paths and means of momentary assessed variables. Separate models were run with and without the means of the momentary assessed variables to determine the relative contributions of average levels and dynamic associations in predicting traditional dispositional scales. Table 3 contains the baseline scale names, R² for each variable, and the error penalization value (i.e., lambda). Supplementary materials provide the specific significant predictors for each model. Using only the GIMME estimated paths, large R² values were achieved for Difficulty Engaging in Goal Directed Behavior, General Interpersonal Distress, Sensation Seeking, and Dyadic Satisfaction in the current romantic relationship. More modest R² were found for Impulse Control Issues, Negative Urgency, Emotional Strategies, Emotional Clarity, Dyadic Cohesion and Consensus. No variance was explained in the remaining scales. Incorporating averages of momentary scales as predictors lead to marked increases in many cases (e.g., Impulsivity, General Distress, Lack of Premeditation, and Negative Urgency).

Table 3.

Proportion of variance accounted for (R²) in baseline scales from GIMME estimated path coefficients and means of momentary scales using adaptive LASSO.

Baseline Measure	With Means		Without Means
	R2	λ	R2	λ
Affect Dysregulation
Nonacceptance	.00	.27	.00	.27
Goals	.33	.17	.33	.17
Impulsivity	.55	.09	.02	.25
Awareness	.00	.23	.00	.23
Strategies	.04	.25	.02	.24
Clarity	.02	.28	.02	.28
Interpersonal Problems
General Distress	.53	.06	.17	.10
Dominance	.00	.13	.00	.13
Affiliation	.00	.09	.00	.09
Impulsivity
Negative Urgency	.12	.18	.04	.20
Lack of Premeditation	.19	.17	.00	.24
Lack of Perseverance	.00	.17	.00	.17
Sensation Seeking	.16	.16	.17	.15
Positive Urgency	.00	.14	.00	.14
Dyadic Adjustment
Dyadic Consensus	.15	.18	.07	.22
Dyadic Satisfaction	.24	.06	.26	.05
Dyadic Cohesion	.08	.12	.08	.12
Affective Expression	.00	.15	.00	.15

Note: λ = adaptive LASSO error penalization parameter.

Network Density

Individuals in the borderline group had an average degree of 10.21 (SD = 4.30, range = 4–22), while those without had an average degree of 9.02 (SD = 2.97, range = 5–18). The two did not differ, t(76) = −1.41, p = .16. In other words, there did not tend to be more significant associations among variables in individual networks for those who met threshold for BPD diagnosis relative to those who did not. We also looked at the extent to which dimensional scores of BPD symptoms were associated with degree. Total number of symptoms was not associated with in-network degree in linear regression, ß = .11, 95% C.I. = (−17 – .47).

Exemplar Models

To more richly illustrate the idiographic nature of these dynamic processes, we describe two individuals from the BPD subsample who had markedly different models, despite sharing a diagnosis. The personalized models are depicted in Figure 3.

Figure 3.

Idiographic models from group iterative multiple model estimation for two exemplar cases. Solid lines represent contemporaneous effects. Dashed lines represent lagged effects. Positive (hot) effects are in red, negative (cold) effects in blue. Strength of effect is indicated by line thickness.

For Participant A, we see strong (i.e., heavy) positive contemporaneous paths among the negative affect scales during interpersonal situations. Affect in social situations is largely divorced from interpersonal behavior of self and other, with the exception of hostility, which is associated with perceptions of lower warmth and behaving less warmly. Perceiving others as dominant is associated with stronger feelings of hostility in the moment. Unlike the affect scales, the interpersonal scales are unrelated, perhaps suggesting that this individual’s behavior is not contingent on situational cues. A few lagged effects emerged, such that feelings of hostility in one situation predicted behaving more dominantly in the next, indicating that affect in one situation tended to influence future interpersonal interactions. The participant’s own interpersonal behavior exhibited a feedback loop, such that increased warmth in one situation predicted increased dominance in the next, but that dominance then predicted decreased warmth.

The interpersonal and affective dynamics of Participant B were starkly different. Many more paths linked affect and interpersonal behavior, and interpersonal perceptions and their own interpersonal behavior were more densely linked. The interpersonal behaviors were linked by contemporaneous paths, such that the participant’s own dominant behavior was associated with lower warmth and higher dominance in the other. This particular positive link between self and other dominance runs counter to interpersonal theory’s complementarity principle, and its violation may be indicative of maladaptivity (Hopwood et al., 2019). Perceiving someone as dominant also was linked with hostility. Several interesting lagged effects emerged, for instance behaving dominantly predicted more positive affect and less fear in the next situation. Experiencing fear predicted behaving more warmly in the next situation, perhaps as the participant sought out support. Although a number of other paths remain to be interpreted, we end by highlighting the positive association between time between interactions (i.e., lag) and fear, such that longer this participant goes between interactions, the more fear they are likely to experience in the next interaction they have, perhaps representing an avoidance process.

Discussion

The prevailing nomothetic research paradigm glosses over important differences in how psychopathology is expressed within patients over time, which may best be ascertained using idiographic research (Fisher, 2015; Wright, 2011). We believe an important next step towards a fuller understanding of psychopathology is the integration of group-level (i.e., nomothetic) and person-specific (i.e., idiographic) modeling techniques (Beltz et al., 2016; Wright et al., 2019). Conceptualizing psychopathology as ensembles of contextualized dynamic processes is particularly relevant for complex disorders like BPD that are defined by cross-domain (i.e., affective, interpersonal, cognitive) interactions (Wright, 2011; Wright et al., 2016). We evaluated the heterogeneity in person-specific structures of interpersonal and affective processes of individuals with BPD diagnoses, compared these to a larger and more heterogeneous sample, linked individual differences in within-person processes to dispositional measures of transdiagnostic dimensions, and in so doing illustrated the utility and feasibility of estimating personalized models of within and between diagnosis processes.

Overall, our results offer some support for the notion that those with the same diagnosis (i.e., BPD) are more similar in their processes than those within a mixed diagnosis sample. However, those within the BPD group also showed substantial heterogeneity within and between functional domains; no two patients were exactly the same. This emphasizes the potential value of building nosologies from the ground up, using individualized models to find commonalities in specific processes among individuals rather than imposing a top-down structure that may not fit any given patient (Fisher, 2015; Wright, 2011). These findings will likely be unsurprising to the frontline clinician: they are well aware that each patient is unique. However, our results do not suggest that the current diagnostic system has not identified important phenotypic manifestations, only that shifting the way we organize our structure of psychopathology could be improved by focusing on the person rather than syndromes of questionable specificity. Estimating individual models and searching for patterns of dynamic processes should identify important processes that may or may not adhere to extant diagnostic syndromes. For instance, using large heterogeneous samples, commonly expressed patterns may reflect generic maladaptive features non-specifically indicative of dysfunction (e.g., group paths; like stress predicts negative affect, as in Wright et al., 2019), whereas rare patterns may isolate higher-value treatment targets that maintain an individual’s (or relatively more homogenous subgroups) pathology.

Consistent with the growing interest in using methods such as GIMME to study multi-domain pathologies such as BPD as ensembles of contextualized dynamic processes (Dotterer et al., in press; Ellison et al., in press; Lane et al., 2019, Wright et al., 2019), we adopted the novel approach of assessing relevant behavior in the key contexts in which the pathology of interest manifests. Given the importance of the interpersonal domain for BPD, we used event-contingent ambulatory assessment reports that were situated in social interactions. As such, we sampled interpersonal and affective information, maximizing our ability to find dynamic patterns of associations within and between these two relevant domains. Our results indicated that the associations within and between both domains showed remarkable heterogeneity across individuals, and that contemporaneous and lagged cross-domain associations existed both for those with BPD diagnoses and those without in the larger mixed sample. This again argues for a personalized-model approach to building nomothetic models of psychopathology rather than relying exclusively on legacy syndromes (Fisher, 2015).

Although not a goal of the current study, our results inform the debate about whether BPD is best understood as an affective disorder (e.g., Linehan [1993] who emphasizes emotion dysregulation) or an interpersonal disorder (e.g., Kernberg and Caligor [2005] who emphasize mental representations of self, other, and self in relation to other). Our results within the BPD subsample suggest that BPD patients are heterogeneous, with some better described by affective dynamics and others better characterized by relations between the interpersonal and affective domains. Thus, it is conceivable that different theorists have simply emphasized different clinical presentations in their descriptions and conceptual models. The question then shifts from which theory is right to which theory fits which patient’s processes (e.g., Hofmann & Hayes, 2019; McWilliams, 2011)? We hasten to add that these results are based on a relatively small sample, and as such represent a compelling albeit preliminary demonstration, but call for replication and extension in larger samples that with protocols that were explicitly designed for personalized models.

GIMME holds considerable power as a data driven approach to develop scalable person-specific models in large samples. However, given the exploratory and data-driven nature of these GIMME analyses, we next used LASSO regression to link paths extracted from GIMME to existing cross-sectional measures of transdiagnostic processes relevant to BPD. The use of the LASSO-based approach, like the current study in general, was meant to be exploratory and hypothesis-generating, rather than confirmatory. The approach’s use of regularization to protect against Type 1 error was in keeping with this strategy. Because of this, we are hesitant to draw strong conclusions about any specific finding, and instead draw the reader’s attention to the fact that relatively strong prediction of dispositional scales was achievable in a number of cases. This is all the more impressive when considering the fact that some dispositional scales were unrelated to content in our ambulatory assessments (e.g., impulsivity), and cross-method prediction is notoriously difficult. The power of context is further demonstrated in the strong dyadic adjustment scale predictions, because many of the reported interactions included romantic partners. However, no prediction was achieved for several of the selected scales.

It is also important to note that much stronger prediction was achievable when using an individual’s average of momentary scales (e.g., average interpersonal warmth, sadness) alongside the dynamic associations among variables. This has implications for how these assessments may best be used in clinical practice. Many clinical researchers have proposed the use of ambulatory assessment and personalized models for “precision” assessment in clinical practice (e.g., Roche et al., 2014; van Os et al., 2013; Wright & Zimmermann, in press). The emphasis in these calls has been on the ability to identify and target the dynamic links among variables, which are often the focus of clinical theories of onset and maintenance of psychopathology (e.g., positive/negative reinforcement cycles; affect regulation), and are the focus of therapeutic interventions (e.g., disassociating stimulus and maladaptive response). The results of such an assessment might resemble the diagrams presented in Figure 3, and could be interpreted by a clinician or presented for collaborative discussion with a patient. Indeed, Fisher and colleagues (2019) have starting using similar assessments to influence treatment planning. Based on the results here, although the unique value of this approach may be in assessing the dynamic links between behavior, it will be important to continue to attend to average levels of relevant behavior and experience. We suggest that this be added to the developing list of considerations requiring attention when translating these methods from bench-to-bedside (Bos et al., in press; Zimmermann et al., 2019).

Another methodological innovation was our use of exogenous variables within the GIMME model to account for various effects of time (Arizmendi et al., 2019). We included the distance in time (i.e., lag length) from the previous rating and raw time passed since the first assessment was completed. Although the latter was included primarily to remove the need for detrending prior to GIMME analysis, the lag length variable allowed us to examine the extent to which time between assessments had effects on any of the affective or interpersonal variables of interest. Future studies could move beyond simply including lag length as a covariate, and incorporate it as a moderator of other effects. Additional variables that might be conceptualized as exogenous (e.g., environmental stressors) might be fruitfully included as exogenous in this fashion.

We believe the interpretations related to heterogeneity in processes are justified, though there are alternative possibilities related to modeling artifacts that must be acknowledged. First, although low numbers of group-level paths were taken as evidence of heterogeneity, they may be driven by low power (Epskamp et al., 2017), especially among participants at the lower end of person-specific sample sizes. Second, observed heterogeneity in paths may be driven by false discovery rates (i.e., type 1 error). The GIMME algorithm is quite conservative, but with eight variables and up to 112 possible paths, even using alpha = .01 as a criterion allows for the possibility of one erroneous path per participant even if all shared the same underlying true model. To the extent this is occurring, this can accumulate to a large number of false associations across large numbers of participants. Finally, recently there has been concern expressed about the replicability of network models in psychopathology research (e.g., Forbes et al., 2017a, 2017b, in press; van Borkulo et al., 2016). To briefly summarize a technical (and hotly debated) literature, it has been observed that specific edges do not replicate in cross-sectional networks when comparing samples that could arguably be assumed to have the same underlying structure, either due to high similarity or very closely repeated assessments. This raises the question whether the observed heterogeneity here might not merely reflect non-replicability driven by the same processes in the cross-sectional data despite an underlying shared structure across individuals. In other words, should the logic not be the same across modeling contexts? We believe that context matters a great deal, and that it is defensible to shift expectations going from highly similar samples of individuals to observations within unique individuals. In the former, we believe it is reasonable to expect high consistency of resulting paths, and in the latter, we believe it is reasonable to expect heterogeneity given that individuals differ so much in so many ways, and each individual generates each sample being compared. With that said, we believe it is important to give serious consideration to what degree issues of measurement and modeling are generating differences in participant structures, and much more basic psychometric work on these issues is needed (Wright & Zimmermann, in press). Finally, GIMME estimates directed contemporaneous paths which, for a variety of reasons, are probably best interpreted as associative but not causal (i.e., place confidence in the association but not the direction of the effect).

As psychopathology researchers recognize the importance of idiographic processes alongside nomothetic processes, new types of research questions will emerge. In the current study, we assessed only interpersonal and affective domains, but there are clearly many other domains of functioning that are relevant to psychopathology research (e.g., cognitive, physiological). Different forms of psychopathology will guide variable selection and sampling. For instance, studies focusing on internalizing pathology would likely want to target negative affect, but also ruminative thought, avoidance, and activity levels (e.g., Fisher et al., 2017). Moreover, in the current study we sampled social interactions, because interpersonal behavior is an important context for the expression of many forms of psychopathology, but especially BPD. Therefore, our interpretations are tied to interpersonal situations, such that contemporaneous associations are those that capture within-interaction processes, and lagged associations reflect spill over from one interaction to the next. Other sampling strategies (e.g., random signals, fixed interval) may be better suited for other forms of pathology or specific research questions (Wright & Zimmermann, in press), and the interpretation of certain parameters might change accordingly.

Larger datasets would provide more power, thus enabling a more fine-grained search for shared and unique associations between more than two domains, providing a more comprehensive picture of psychopathology. Adding participants increases generalizability and the power to examine associations and make predictions outside of the model. However, power within the individual models that GIMME estimates is not related to adding more participants, but rather including more observations per participant. In addition, future investigations should look to identify paths that are particularly good markers for psychopathology. It is likely that there will be many paths that are linked to what we think of as traditionally similar diagnoses; however, thinking in terms of the maladaptive dynamic process is likely conceptually closer to the phenomena of interest for psychopathology researchers and more relevant to how clinicians approach treatment. Investigators interested in particular theoretical models of psychopathology can test the extent to which individuals are similar in the dynamic processes proposed by the theory. This can be extended into treatment research: Does a patient’s deviation from the processes proposed by a treatment model reduce the intervention’s effectiveness?

The current study had several limitations. First, our sample size was relatively small. Although our study entailed secondary analyses of an existing data set, and thus meant to be exploratory (consistent with a proof-of-concept) rather than confirmatory, it is possible that larger samples may have detected more associations with baseline data. Second, we had to exclude a majority of the parent sample due to insufficient interaction-based data. These exclusions left us with 24 participants in our BPD subsample, and 54 additional participants in our mixed diagnosis subsample (78 total), these sample sizes compare favorably to existing studies and were adequate for the purposes of our exploratory aims. Each individual’s number of observations is a consideration for all person-specific research. Just as between-person sample size dictates power and precision in cross-sectional research, a sufficient number of assessments per participant is needed in process-oriented research. Relatedly, because idiographic research is specific to individuals, who is included in studies is a primary consideration. Because the current study used participants that were recruited for a study of romantic couples and estimated processes within social interactions, the interpretation of models and paths should be constrained to this population and context. Some few diagnoses, bipolar and frank thought disorder, were excluded. Effects can appear contemporaneous when assessments occur at a slower rate than the change in the measured process (Granger, 1969). It is possible, even likely, that at least some of the contemporaneous effects GIMME identified actually occur over time rather than simultaneously. This is a limitation is shared by all EMA collection that does not use continuous assessment (e.g., physiological data). Finally, although we used LASSO regression to link the many parameters generated by GIMME with traditional cross-sectionally assessed scales, other techniques such as adaptive LASSO or elastic net may provide better characteristics for the aim of linking the two. Ideally these would incorporate cross-validation or hold out samples as well, which we did not do here.

Clinical researchers are shifting emphasis from studying heterogeneous clinical syndromes to identifying and investigating transdiagnostic features of psychopathology (e.g., Kotov et al., 2017; Wright & Simms, 2015). While this is promising, the reigning nomothetic paradigm of psychopathology research often prioritizes interindividual differences over intraindividual processes. Here we argue that psychopathology is best understood as contextualized dynamic processes, and these manifest within an individual in a complex system over time and circumstances. The current study adopted this perspective in studying dynamic affective and interpersonal processes in social situations in a sample with a range of pathology and in a subsample whose members all shared a BPD diagnosis. We showed that the structure of each individual’s processes was unique with some evidence for shared processes across individuals, particularly within the BPD group. The use of GIMME and similar idiographic techniques, when applied to large samples of individuals, provide the necessary tools to develop generalizable models of psychopathological processes that respect the individual as a whole system. The goal is to move beyond identifying who has maladaptive functioning in a given affective, behavioral, or cognitive domain, but understanding when, in which context, and how dysfunction in each of these domains interfaces with the others. We believe this will unlock new insights and accelerate our mechanistic understanding of what leads to the development, maintenance, and resolution of psychopathology.

Public Health Significance Statement:

This study highlights the importance of understanding psychopathology using individualized models. We demonstrate research techniques that can be used to develop personalized models of psychopathology as well as search for commonalities across individuals.