Trait-State-Occasion Model of Borderline Personality

Figure 1 illustrates how the TSO model was used to parse borderline PD variance into 1 stable (i.e., time-invariant) trait factor and 5 time-varying occasion factors. The trait factor reflects the stable core of borderline proneness that does not fluctuate over the follow-up interval, whereas the occasion factors represent causes of borderline PD that vary over time. Autoregressive pathways were specified between contiguous occasion factors to capture the continuity of time-varying contributions to borderline PD. The state factor variances were constrained to 0, reflecting the fact that borderline PD variation at any point in time was completely partitioned into stable and time-varying components. Following Cole et al.’s (2005) recommendations, we (i) constrained factor loadings of borderline PD indicators to equality over time; (ii) constrained the residual terms of the latter 4 occasion factors to equality; (iii) fixed the 4 autoregressive parameters to equality, except the last autoregressive path, which spanned a 4-year (as opposed to 2-year) interval; and (iv) allowed residual correlations among the identical borderline PD indicators across waves.

Table 1 presents the TSO model parameter estimates. The model offered a good fit to the data, χ²(887) = 1,225.47, p < .001; CFI = .98; TLI = .98; RMSEA = 0.02; WRMR = 1.12. The trait factor variance represented 43% to 45% of the total borderline PD variability across waves. This proportion was not constant across follow-up assessments because borderline PD variability changed over time. For instance, the trait factor contribution was largest at year 10 because model-estimated variability in the borderline PD construct was smallest at that stage. Occasion factor variance was larger than 0 at all time points (ps < .001), and moderate continuity was observed across adjacent occasion factors (standardized coefficient range: .55−.59).²

We compared the size of the time-invariant and time-varying factors by computing a chi-square difference test of the ratios of time-invariant to total variance and time-varying to total variance using the model test command in Mplus. This test was not statistically significant, χ²(1) = 0.81, p = .37, thus providing no evidence of a difference in magnitude between the two longitudinal factors. Further, a test of whether the ratio of time-invariant to time-varying variance was different from unity was also non-significant, χ²(1) = 1.06, p = .30, indicating that the ratio was not significantly different from 50/50. ³

A Hybrid Model of Borderline Personality Disorder Change

We compared the time-invariant components of putatively temperamental versus acute borderline PD criteria, as distinguished by Zanarini et al. (2007). The temperamental symptoms, which are theoretically more enduring, include impulsivity, unstable relationships, chronic emptiness, and intense anger. The theoretically acute symptoms include abandonment fears, suicidal behavior and self-harm, affective instability, identity disturbance, and stress-related paranoia. We estimated two separate TSO models in which the state factors were defined by either the 4 temperamental diagnostic criteria or the 5 acute diagnostic criteria. Table 2 presents model fit indices and parameter estimates for both models. The TSO model fit the temperamental symptom data well, and the percentage of trait variance (to total variance) ranged from 35–43% across waves. The corresponding analysis for the acute symptoms also produced acceptable model fit, and the ratio of trait to total variance in this model was 49–51%. Thus, the proportion of variance in borderline pathology that was stable across 10 years was decidedly not higher for putatively temperamental symptoms, relative to acute symptoms. These results fail to support that the hybrid model proposed by Zanarini et al., who found that impulsivity, unstable relationships, chronic emptiness, and intense anger showed slower mean-level declines in the MSAD dataset than other diagnostic criteria. This hybrid model evidently does not extend to the differential stability (as opposed to mean-level, or absolute, stability) of borderline PD features.

Trait-State-Occasion Models of Big Five Traits

As a preliminary step toward estimating the association between the time-invariant factors for FFM domains and borderline PD, we tested the fit of TSO models for each FFM domain separately. There was evidence of longitudinal measurement invariance for all FFM traits (results available in the online supplement). All five TSO models had interpretable patterns of factor loadings and provided a good fit to the data. Tables S1–S5 in the online supplemental material report the fit indices and parameter estimates from the univariate TSOs for each FFM domain. The mean (across waves) proportion of variance accounted for by the time-invariant factors was 47.6%, 66.6%, 74.4%, 77.5%, and 86.1% for Neuroticism, Conscientiousness, Agreeableness, Extraversion, and Openness, respectively.

Next, we estimated bivariate TSO models for each combination of borderline PD and the FFM domains. This set of models allowed us to estimate the longitudinal correlations between the time-invariant and -varying factors of all pairs of constructs. Each of the bivariate models provided a good fit to the data (full results available upon request). The lower triangle of Table 3 displays the time-invariant factor correlations derived from these analyses. Borderline proneness was most closely linked to Neuroticism (r = .74) and most weakly associated with Openness (r = −.15). Correlations with the other FFM domains were more moderate (r range: −.34 to −.48). Also, the upper triangle of Table 3 presents the residual correlations, which were constrained to equality over follow-up, among the time-varying factors from each bivariate model. These values reflect the association among transient (i.e., not perfectly stable) influences on the personality constructs. The pattern of FFM time-varying factor associations with borderline PD very closely mirrored the time-invariant factor correlation results. That is, the dynamic element of borderline PD was strongly related to the dynamic element of Neuroticism (r = .64); less robustly with Conscientiousness, Extraversion, and Agreeableness (r range: −.30 to −.32); and minimally with Openness (r = .22).

 

 

 

 

 

 

 

 

Note. N = 662. Correlations reflect associations between factors extracted in “bivariate” trait-state-occasion models. Time-invariant (i.e., trait) factor correlations appear below the diagonal, and time-varying (i.e., occasion) residual factor correlations above the diagonal. For example, in a joint trait-state-occasion model of borderline personality disorder and Neuroticism, the correlation between the time-invariant factors—reflecting the perfectly stable components of these two constructs across the study timeframe—was .737. Thus, 15 bivariate trait-state-occasion models were estimated to compute the correlations presented here.

^ap < .001,

^bp < .05.

We repeated this process for the FFM composite of borderline PD that is defined by the Neuroticism, Openness, and Conscientiousness facets identified by Lynam & Widiger (2001). In testing the longitudinal measurement model for this construct, we found that, whereas the model fit acceptably overall, χ²(678) = 1,373.42, p < .001; CFI = .95; TLI = .94; RMSEA = 0.04; SRMR = 0.08, the two Openness facets had small factor loadings on the latent trait. The Feelings facet loading range was .10 to .13 (ps < .001) and the Actions loading range was −.20 to −.23 (ps < .001), suggesting they were not operating as indices of borderline trait vulnerability as expected. We therefore dropped these two indicators (cf. Warner et al., 2004), and the resulting measurement model fit improved, χ²(355) = 728.86, p < .001; CFI = .97; TLI = .96; RMSEA = 0.04; SRMR = 0.07. The univariate TSO model of this FFM composite fit the data acceptably, and revealed that 48% to 56% of variation was completely stable over time (see Table S6 for fit statistics and parameter estimates).

Finally, a bivariate TSO involving borderline PD (as reflected in DIPD-IV symptoms) and the FFM borderline trait composite was estimated to evaluate the correspondence of the disorder and trait constructs. This TSO fit the data adequately (χ²(2,599) = 3760.86, p < .001; CFI = .93; TLI = .92; RMSEA = 0.03; WRMR = 1.30) and revealed a time-invariant factor correlation of .81 (p < .001). The correlation between the time-varying factors of borderline PD and the FFM composite was smaller, yet still strong (r = .63, p < .001). Overall, these results suggest that the majority of between-person variation in borderline proneness—both the stable and dynamic components—can be accounted for by selected FFM traits.

Construct Validity of Borderline Proneness

This study builds upon the large literature connecting borderline PD to basic traits (Samuel & Widiger, 2008) by showing that normative personality traits could explain a substantial proportion of between-person variance in borderline PD features. Our approach specifically extends previous cross-sectional empirical work on observed variables by illuminating the “true” correlations—typically obscured to some extent by measurement error, mood-state distortion, and other transient sources of noise—between the stable components of these trait-like constructs. Neuroticism (r = .74), Agreeableness (r = −.48), and Conscientiousness (r = −.38) had particularly robust relations with borderline proneness. The relations between these stable components map on fairly closely with the genetic correlations (i.e., associations among completely fixed genetic contributions) among these same constructs in a population-based Norwegian twin sample (rs of .71, −.44, and −.58 with borderline PD symptoms, respectively) (Czajkowski et al., in press). Moreover, we showed that the dynamic (i.e., state-like) elements of borderline PD were fairly well-captured by the dynamic elements of normative personality across follow-up waves in this study. The pattern of time-varying factor correlations largely mirrored the nature of associations among the stable, time-invariant factors, and ranged from r = .22 (Openness) to r = .64 (Neuroticism).

Second, we evaluated the association of borderline proneness with the trait-like component of a FFM facet-level composite of borderline PD dimensions based on expert consensus. This constellation of lower-order traits was originally theorized to represent the core features of borderline pathology at a finer level of resolution than the FFM domains permit (Lynam & Widiger, 2001). After omitting two facets from the Openness domain that did not empirically align with the other FFM indices of borderline PD in our data, we discovered that the time-invariant factors of borderline PD (based on clinical interviews) and the FFM composite were very strongly related (r = .81), as were the corresponding time-varying factors (r = .63). This lends further validation to the hypothesis that borderline PD features can be captured systematically—albeit not completely—by normative personality variation. We argue that future efforts to trace the stable personality underpinnings of borderline and other PDs—whether through a FFM lens, or other models of personality structure—should consider parsing the fixed substrates of PD from transient variation likely to obscure the time-invariant associations of disorder constructs with external factors (cf. Naragon-Gainey, Gallagher, & Brown, 2013).

We caution that while the FFM offers good statistical prediction of borderline proneness in these data, the FFM constructs may not be the only, or even the most important, foundations of borderline PD. Recent findings on the multivariate structure of PDs suggest that borderline PD overlaps markedly with a general liability for PD (Sharp et al., 2015; Wright, Hopwood, Morey, & Skodol, 2016). That is, in bifactor models that extract a general factor reflecting the common substrate of all PDs, borderline PD has minimal overlap with other PDs independent of this general liability (e.g., Wright et al., 2016). These results have been interpreted as indicating that borderline PD largely reflects core deficits in interpersonal functioning and identity formation, which form the basis of the general definition of PD in emerging PD diagnostic models (e.g., Bender et al., 2011). Borderline proneness may therefore represent, at least in part, core adaptive functioning problems that are common to the full range of PDs (see Kernberg, 1984; Turkheimer, Ford, & Oltmanns, 2008).

We looked beyond the FFM to further explicate the construct validity of the stable and dynamic borderline PD factors. All three forms of childhood maltreatment investigated here had strong effects on stable borderline proneness and virtually no influence on more acute fluctuations in symptom severity. This finding is consistent with epidemiological evidence that childhood adversities instigate sustained risk for common mental disorders (e.g., McLaughlin et al., 2010). Paralleling the maltreatment results, we observed negative temperament and positive temperament had much more potent effects on borderline proneness than transient borderline pathology. Consistent with the correlation between neuroticism and borderline PD TSO trait factors (r = .74), negative temperament had a particularly large influence on borderline proneness (r = .79). Contradicting hypotheses, the disinhibition dimension appeared to be twice as strongly related to the dynamic component as to the stable component of borderline pathology. While disinhibition is both theoretically and empirically linked to borderline PD (e.g., Trull, 2001), no prior work has parsed its associations with enduring versus transient elements of the disorder. It is possible that in this patient sample disinhibition was particularly related to acute symptoms that motivated treatment. Often, such symptoms (e.g., self-damaging impulsive behaviors, non-suicidal self-injury, quasi-psychotic thinking) are the proximal triggers of clinical attention. We note, however, that this effect was unexpected and could be specific to treatment-seeking groups. More work in a variety of longitudinal research designs is needed to follow up on this finding and the other criterion validity results.

The GAF result was also partially inconsistent with our hypotheses. GAF ratings were moderately inversely related to both time-invariant and -varying factors, whereas we anticipated a specific linkage with dynamic borderline features (i.e., we envisioned that symptom flare-ups would be accompanied by worsening psychosocial impairment). Our post hoc interpretation of this result is that impairment is, to some extent, a stable feature of PD patients. In fact, multiple longitudinal CLPS studies have converged on the finding that, even in the context of PD symptom improvement, psychosocial functioning remains fairly steady over time, especially for borderline PD patients (e.g., Gunderson, Stout et al., 2011). Therefore, although GAF ratings are intended to capture patients’ current status, we argue that empirically they probably also reflect a somewhat enduring property of the patient.

We believe that, taken together, these associations with external criteria support the construct validity of the stable and dynamic elements of borderline PD. This study provides some evidence that borderline proneness is associated in expected ways with events and processes (e.g., child abuse, negative temperament) that are often implicated in etiological theories of borderline PD. Moreover, many of these associations were undeniably robust and large in magnitude, probably because borderline proneness excludes all transient sources of borderline PD variation (e.g., measurement error, short-lived environmental inputs) that are theoretically unrelated to most prominent etiological factors. At the same time, the dynamic elements of borderline pathology were systematically related to external correlates. For instance, time-varying borderline PD features were consistently associated with yearly fluctuations in normative personality traits and interview-based GAF ratings at study baseline. We advise investigators to prioritize research designs and statistical plans that can delimit the component(s) of borderline PD that is best suited to the particular research setting. For instance, studies on the genetic underpinnings of disorder would likely concentrate on borderline proneness, whereas research on psychological treatment or acute stressors may align better with dynamic components of borderline pathology.

Limitations

Our study benefitted from psychometrically sound interview assessments of borderline PD symptoms and self-report assessment of basic traits, a large patient sample with diverse principal PD diagnoses, and a decade-long follow-up data structure. A number of study limitations should be considered alongside these design strengths, however. First, the longitudinal structure of borderline pathology depends on several study attributes, including timeframe, sample illness severity, assessment method, and developmental stage. For instance, we might expect greater instability over longer intervals, in patient (as compared to non-clinical) samples, and in young people (Roberts, Walton, & Viechtbauer, 2006; see also Wright, Zalewski, Hallquist, Hipwell, & Stepp, 2016). Along those lines, we found that Neuroticism levels were highly variable over 10 years in CLPS (i.e., as for borderline PD, nearly half of variation over time was state-like), whereas other research has reported stronger consistency, albeit over shorter intervals (e.g., Prenoveau et al., 2011). We recommend additional research into the longitudinal structure of Neuroticism and other FFM domains in diverse samples to determine the factors (e.g., illness severity, psychiatric treatment) that influence personality stability.

Second, multiple sources of information about patients’ personality disorder symptoms could strengthen longitudinal models of borderline PD (e.g., Cole et al., 2017). TSO modeling is equipped to isolate the shared variance among patient, interviewer, and informant reports—which often disagree (Hopwood et al., 2008)—of PD symptoms. Third, the FFM trait composite of borderline PD overlapped significantly with Neuroticism, such that it may not offer much incremental explanation of borderline PD. This was partly because the Openness facets originally theorized to relate to borderline PD were not empirically related to the FFM composite in this dataset, and partly because borderline PD is intimately intertwined with Neuroticism (e.g., Trull, 2001). Other personality disorders are represented by a greater variety of FFM facets (Lynam & Widiger, 2001). Nevertheless, our results here regarding Openness facets suggest the need for further refinement of the personality architecture of borderline PD (e.g., DeShong et al., 2015). Fourth, the borderline PD diagnosis is notoriously heterogeneous (e.g., Clark, 2007), and its basic building blocks (e.g., affective disturbance, attachment insecurity) may vary in terms of stability. The TSO framework does not permit symptom-level (i.e., single indicator) analyses in this dataset, but we recommend future research into the temporal course of intermediate phenotypes for borderline PD. The FFM domains arguably represent one model of borderline PD heterogeneity (e.g., Widiger & McCabe, 2018), and our results indicate that they are differentially stable over time. Fifth, there was significant attrition over follow-up in CLPS (although dropout rates compared favorably to other long-term studies). We used a missing data procedure that makes less restrictive assumptions than most other methods (Widaman, 2006), but some possibility for bias in our parameter estimates remains. Finally, there is considerable debate about the nature of the borderline PD concept itself, with some conceptualizations situating it as a general factor of personality pathology, others as a specific syndrome, and others as a superfluous concept. Thus, results may differ depending upon how borderline PD is conceptualized and measured.