Categorical unit of analysis

An early two-wave (test–retest) study of the stability of BPD in a group of 65 inpatients, using the DIB, found that 60% (39 of the 65) maintained their BPD diagnosis at 2-year follow-up (i.e., met criteria of score of 7 or more on DIB), while the remaining 40% (26 of 65) dropped below the diagnostic criterion level (Links, Mitton, & Steiner, 1992). This portion of the sample (n = 26) may still have displayed nontrivial borderline pathology, but they had dropped below the threshold for categorical diagnosis. According to Links et al. (1992), positive response to DIB questions regarding “unstable work or school history during the past 2 years,” “self-mutilation,” and “serious pattern of substance abuse,” as well as the presence of a lifetime, comorbid affective disorder and substance abuse were significantly higher in those with a stable BPD diagnosis over the 2 years of the study compared to those who dropped below threshold. Higher rates of physical abuse during childhood, history of unemployment during the 5 years prior to the study, and higher rates of psychiatric illnesses in first-degree relatives were also found in this continuous BPD group compared to the remitting subsample.

Categorical and dimensional units of analysis

A naturalistic, multiwave investigation of inpatients and outpatients known as the Collaborative Longitudinal Personality Disorders Study (CLPS; Gunderson et al., 2000) provides useful data on the short-term stability of BPD pathology. In this study, begun in 1996, the investigators examined four PDs (borderline, avoidant, schizotypal, and obsessive–compulsive PDs) over a 12-month period, with assessments at baseline, 6 months, and 12 months (Shea et al., 2002). Only 41% of those initially diagnosed with BPD at the beginning of the CLPS maintained this diagnosis at 1-year follow-up (Shea et al., 2002). When Shea et al. (2002) examined BPD symptoms from a dimensional perspective, the decrease in symptoms was found to be greatest during the first 6 months of the study, with minimal change during the second 6 months, which was a pattern reflected in the other three PDs as well. Individual differences in BPD features remained stable over time, which suggested that subjects maintained their position in the rank order with respect to BPD features over time.

The data from the CLPS, however, present complexities in terms of interpretation owing to design limitations of the study. Notable among these limitations is the fact that subjects in the study were all drawn from treatment settings, thus possible change due to time cannot be disentangled from change due to treatment in these data (i.e., no subjects were studied who had never been in treatment and many subjects in the study remained in treatment). Furthermore, the diagnosticians in the CLPS study were not blind to the prior assessments conducted on the patients they interviewed, with many subjects being reinterviewed by the same diagnostician. For studies using interview-based assessments, a critical design feature is the use of blinded assessment such that the same subject is never seen by the same interviewer more than once. This refinement helps to ensure against “halo effects,” which diminish validity and artifactually elevate stability estimates (see Loranger et al., 1991). It is well known in the assessment literature that memory of information about a prior assessment impacts adversely the validity of a current assessment (Nunnally & Bernstein, 1994). Thus, any study that seeks to inform the scientific issue of stability of BPD or any other PD cannot do so effectively without the use of multiple blinded assessment across all study waves. Finally, the BPD, and other PD, subjects included in the CLPS were selected on the same measure (independent variable) that would then be used later to study stability or change (dependent variable), namely, Diagnostic Interview for Personality Disorders (DIPD-IV; Zanarini, Frankenburg, Sickel, & Young, 1996) assessed BPD symptoms, which raises concerns about endogeneity in interpreting the results. The absence of selection independence (i.e., a different instrument is used for subject selection than the instrument used subsequently to track changes) also creates a situation where regression to the mean effects could be pronounced.

Putting these methodological concerns aside, it is informative to attempt to glean insights from the CLPS data with regard to BPD. A particularly interesting feature of the CLPS data on BPD is that fact that “sudden remissions” of BPD diagnosis and symptoms were observed in this sample (Gunderson et al., 2003). Gunderson et al. (2003) reported on 18 initially BPD diagnosed subjects that were drawn from the subsample of 175 individuals who received a baseline diagnosis of BPD based upon the criteria of the DIPD-IV (Zanarini et al., 1996). As with the CLPS data reviewed above, these subjects were also assessed 6 months, 1 year, and 2 years following baseline. A subgroup of 18 subjects was found to reveal dramatic reductions in borderline psychopathology at 6 months (defined as having met two or fewer of the diagnostic criteria), and maintained this decrease at 1-year follow-up. The investigators further examined this group to assess for the potential roles of four variables, remission of comorbid Axis I disorders, positive changes in negative life situations, effects of therapeutic interventions, and inaccurate baseline diagnosis, in the BPD remissions that were observed. Axis I remission (7 of 18) and relief from stressful life situations (9 of 18) were the two variables found to most frequently be the key factor in the decrease in BPD symptoms. The four variables were ranked with regard to importance for each individual based upon the consensus judgments of the project coordinators of the sites from which the patients came. Although other studies (noted above and below) reveal evidence that BPD symptoms decrease over the course of an individual's life, this subsample from the CLPS study is one of the few to indicate that more rapid and dramatic change can be witnessed in BPD, resembling an acute disorder rather than the chronic condition it has usually been assumed to be. Gunderson et al. (2003) apparently did not, however, compare the rate of occurrence of the four predictive variables in this subsample to rate for the same variables in the remainder of the BPD sample who did not remit or to those who did not meet criteria at 6 months, but relapsed at 1 year. It is therefore unclear whether the life changes mentioned above occurred with the same frequency in these other groups as well. If the rates of occurrence for these four variables did not differ across the remitting BPD and nonremitting BPD subjects, then their predictive importance would be somewhat questionable. Regardless of the importance of these four predictive variables, it remains unclear whether other contributing factors are being overlooked in the 18 cases that displayed spontaneous recovery. For example, there may have been baseline differences in temperament or personality systems that may have predicted change in those subjects who improved dramatically versus those that did not.

Categorical unit of analysis

Clarke, Hafner, and Holme (1995) assessed the stability of BPD diagnosed individuals in a psychiatric inpatient population, which was receiving no consistent therapy except for medication. They found BPD to be a stable diagnosis in 38 of the 43 cases for which they had sufficient data available at 3-year follow-up. These data suggest relatively high stability for the BPD diagnosis; however, this high rate of stability may have been affected by the method of initially selecting candidates. Only those individuals who had presented with obvious BPD symptoms on a multiple number of admissions were included, perhaps biasing the sample for exceptionally severe cases with more persistent BPD features.

In a study by Links, Heslegrave, and van Reekum (1999), an attempt was made to examine whether stability or remission of a BPD diagnosis could be predicted by assessing the role of impulsivity in the disorder. The study subjects were psychiatric inpatients (ages 18–65) from four hospitals and they were diagnosed as definite BPD or non-BPD based on DIB criteria. A 2-year follow-up was conducted only with those receiving an initial diagnosis of BPD, whereas a second follow-up was conducted for all available subjects from the initial 130 subjects, 7 years after initial intake. Based upon the initial intake and 7-year data for subjects diagnosed with BPD at baseline, 52.6% (30 of 57) no longer qualified for a diagnosis of BPD. Impulsivity per se showed high reliability over this time period, and it was found that the impulsivity rating at intake was the best predictor of degree of borderline psychopathology at 7-year follow-up. The group of individuals who had data at all three points (those with an initial diagnosis of BPD) was divided into remitting and persistent BPD subgroups. Changes over time were assessed and compared between the two groups. The remitted group displayed greater decreases in impulsivity during the first 2 years, as well as less severe psychotic symptoms and less disruption of affect and impairment of interpersonal relations at 2- and 7-year follow-up compared with the group with a persistent BPD diagnosis. This study indicates that impulsivity may play a key predictive role in understanding change in borderline PD, and may provide important predictive information regarding the long-term course of the disorder in an individual. It should be noted that the those subjects who dropped out of the Links et al. (1999) study tended to be male. Furthermore, as the subjects in this study were from an inpatient population, it appears that additional investigations would be required to if these findings would generalize to males and outpatient or community populations. Nonetheless, over half of the sample studied no longer qualified for a BPD diagnosis at 7-year follow-up, a pattern of data again suggesting considerable change and flexibility in BPD.

Zanarini, Frankenburg, Hennen, and Silk (2003) reported on a multiwave study that included 290 borderline inpatients and 72 patients with other PDs. The diagnostic instrument used in this study was the Revised DIB (Zanarini, Frankenburg, & Vujanovic, 2002). The rich findings from this longitudinal study are quite striking and provide considerable food for thought. Zanarini et al. (2003) found that 34.5% of borderline patients remitted after 2 years, 49.4% after 4 years, 68.6% at 6-year follow-up, and 73.5% over the entire 6 years of the study (including all subjects with data at 2-year follow-up). Nearly three out of every four BPD patients in this study no longer met the diagnostic threshold for BPD at 6 years. Moreover, recurrence rates (meaning once again meeting criteria after remitting at a previous time period) were low, with 6% recurring at 4- and 6-year follow-up and 12% representing with criteria over entire study. These investigators sought to illuminate this remarkable pattern of change in the BPD patients. They found in an examination of specific symptoms that impulsive symptoms were the first to remit, followed by interpersonal and cognitive difficulties, and followed finally by the affective features. The Zanarini et al. (2003) data provide compelling evidence that BPD is clearly not a traitlike and enduring disorder that is resistant to change. If anything, the Zanarini et al. (2003) provide interesting data to ponder with respect to the true flexibility of BPD, mindful that the subjects were in treatment during the study period.

Dimensional unit of analysis

The loss of data that attends the sole use of categorical measures of PD constructs is well known and the clear-cut superiority of dimensional metrics for the study of growth and change is soundly established in the longitudinal methodology area (Singer & Willett, 2003). There are currently two studies of PD phenomenology that have followed subjects for 3 years or more. Both of these studies have made use of dimensional metrics for the assessment of stability and change.

The first study is the LSPD, which is under the direction of the first author (Lenzenweger, 1999; Lenzenweger et al., 1997; Lenzenweger, Johnson, et al., 2004). The LSPD is the first NIMH-funded prospective, multiwave longitudinal study of PDs, personality, and temperament that was begun in 1991. The LSPD is a naturalistic study that is following two samples of subjects across the life span: one sample was deemed to be at increased risk for a PD at the inception of the study, and the other sample was at diminished risk.

The LSPD established a prevalence rate for the PDs in a nonclinical sample (Lenzenweger et al., 1997) of approximately 11%, with a prevalence of BPD of 1.3%. These figures were subsequently replicated (e.g., Torgersen et al., 2001). In one of the initial reports from the LSPD, the traditional repeated-measures analysis of variance (MANOVA) approach was used to evaluate the stability of mean differences over time for PD ratings on the Personality Disorders Examination (PDE) as well as the symptom dimensions of the Millon Clinical Multiaxial Inventory (MCMI-II; Lenzenweger, 1999). It was found that the PD dimensions, including BPD, tended to decrease over time, and the effect sizes suggested some appreciable amount of change. With respect to BPD, both the PDE and the MCMI-II BPD dimensions revealed evidence of significant declines in symptomatology over time (particularly among the study subjects at increased risk for a PD). Lenzenweger (1999) also examined the stability of individual differences for the Axis II disorders on the PDE and the MCMI-II. The results of these correlational analyses revealed substantial evidence of rank order stability over time over a 4-year time. With respect to BPD, both the PDE and MCMI-II dimensions displayed substantial stability of individual differences over the 4-year period with correlations of .50 and .64, respectively (i.e., Wave I by Wave III associations). The approach taken by Lenzenweger (1999), however, represented a first pass at the issues of stability and change in personality pathology over time. Clearly, the data from the 1999 report suggested there was evidence of stability in the PD features over time, at the same time they suggested some degree of change and flexibility as well. This situation suggested that there was considerable meaning and direction yet to be extracted from these data to shed light on issues of stability and change, which suggested the need for a more fine-grained analysis.

How is it that more information could be extracted from these PD data? The stability of individual differences in the PDs over time that was documented in the Lenzenweger (1999) report was revealed using the between-wave correlation approach. However, this approach only provided thin two-wave “snapshots,” as it were, of the entire spectrum of interindividual differences in growth. The average levels of PD symptomatology, assessed using repeated MANOVA, revealed mean level stability and some evidence of PD feature reduction over time (albeit with small effect sizes). However, the MANOVA approach (a) could not accommodate the existing unequal spacing of PD assessments across study subjects (a typical reality of longitudinal studies); (b) could not disentangle important aspects of individual change such as initial levels of symptomatology and rate of change as typically implemented; and (c) implausibly assumed comparable growth across all subjects, wherein heterogeneity of growth is more likely. As such, although the correlational and MANOVA analyses reported previously represented a reasonable first pass through the LSPD data (Lenzenweger, 1999), they did not tap the true richness therein. An analysis of individual growth trajectories (Singer & Willett, 2003; Willett, 1988; Willett, Singer, & Martin, 1998) was required to represent this richness and to improve the understanding of stability/change in PDs over time.

The analytical technique ideally suited for maximal extraction of information from these LSPD data is IGC analysis (Singer & Willett, 2003; Willett, 1988; Willett et al., 1998). Therefore, Lenzenweger, Johnson, et al. (2004) undertook such an analysis of IGCs for the LSPD subjects using their International PDE (IPDE) dimensional scores as the metric for study over time. Lenzenweger, Johnson, and Willett (2004) reported that personality pathology, when viewed from the IGC perspective, was actually quite flexible. The IGC perspective allowed for the simultaneous analysis of level and rate of change parameters for personality pathology; thus, each subject was uniquely characterized with respect to their level of pathology over time as well as the rate at which that pathology changed. The general trend observed across all PDs was the decline of symptomatology with time. In fact, on the PDE, PD features declined at a rate of 1.4 features per year. When considering BPD features over time, such features were seen to decline significantly over time as well, and the amount of change represented a medium-sized effect. The IGC approach also allowed for the introduction of individual difference variables to predict variability in overall level of pathology and rate of pathology change over time. It was observed that BPD features were higher among those subjects deemed to be at increased risk for a PD, based upon initial selection, and the rate of change (i.e., decline) was also related to this between-subjects variable (i.e., at-risk subjects declined in BPD features at a faster rate over time; Lenzenweger, Johnson, et al., 2004). What was particularly striking about the BPD results from this IGC analysis was that they revealed considerable variance in the growth parameters (level, rate of change [slope]) remaining for explanation. In other words, the initial classification of the subjects (high-risk), gender, and age at entry into the study did not predict all of the variability that was inherent in the data. Lenzenweger, Johnson, et al. (2004) proposed that additional variance in these growth parameters might be explained by model-based predictions from a neurobehavioral model of the PDs (Depue & Lenzenweger, 2001, 2005). Such an analysis is the focus of the second part of this paper.

An additional multiwave naturalistic study that complements the LSPD has been undertaken under the direction of Patricia Cohen at the New York State Psychiatric Institute; it is known as the Children in the Community Study (CCS). The CCS is a community-based longitudinal investigation aimed at the examination of stability and change in PD trait levels between early adolescence and early adulthood. The CCS is a highly creative project that has made particularly effective use of rich archival data on a large number of individuals residing in the community in upstate New York to illuminate the longitudinal nature of PD symptomatology (Johnson et al., 2000). The subjects in the CCS were 14 years old, on average, at the beginning of the study and were reassessed 2 and 8 years later (ages 16 and 22, respectively). Like the LSPD, the CCS has the important methodological feature of community-based subjects, which allows one to study change/stability that is not the result of systematic or directed treatment effects. The general findings for PD features in the CCS are highly consistent with those of the LSPD, namely, PD features were observed to decline over time across the different disorders. Those subjects who presented elevated PD features during adolescence continued to display elevated levels of features during the early adulthood assessment, albeit fewer features compared to the assessment at adolescence. The CCS data also revealed that a considerable portion of change occurred from the Wave I to Wave II assessments, a feature of the data consistent with the LSPD (Lenzenweger, 1999).

With respect specifically to BPD features, Johnson et al. (2000) noted a 59% decline in the number of BPD features over the study period and this decline appeared to be relatively linear in shape. The rank order stability of BPD features in the tended sample was r = .31 over the study period, with ipsative stability for BPD being similar (intraclass correlation [ICC] = .28). Johnson et al. (2000) noted that the linear nature of the decline in BPD traits over time indicates that patterns of change across the various PDs require further exploration to enhance our understanding of the predictor of change. Although this study did not evaluate other potential predictors of change beyond age and initial symptom level, it highlighted the importance of multiwave data and the opportunity that such data represent for discerning predictors of change. These data, when viewed developmentally, are especially important because they document instability of BPD features during the period of adolescence: they clearly suggest a dynamic picture for BPD during this period of growth and development.

Aside from the LSPD and the CCS, there is but one other longitudinal study of BPD features conducted in a multiwave format. Najavits and Gunderson (1995) reported a waxing and waning course for BPD features over time in a small sample (n = 37) of female BPD diagnosed patients. A pattern of overall improvement over time was found in this 3-year naturalistic study. The assessments were conducted at baseline, 1, 2, and 3 years, with borderline symptoms measured using the DIB. Najavits and Gunderson (1995) found that 28 of the 37 subjects improved (20 showed erratic improvement, 8 showed steady improvement), but four subjects did show higher DIB scores at 3 years. It was found that depression and anxiety, as measured by the System Checklist-90, were predictive of poorer outcome at 3 years. Change, however, not unlike that observed in the CLPS study, was confounded with treatment effects; thus, it is unclear to what extent BPD features declined naturally over time versus declines that were in response to directed/systematic treatment.

Subjects

The 258 subjects in the LSPD (Lenzenweger, 1999; Lenzenweger et al., 1997, 2004) were drawn from a population consisting of 2,000 first-year undergraduate students. Subjects were assigned to either a “possible PD” (PPD) or “no PD” (NPD) group according to the IPDE-Screen (IPDE-S; response rate = 84.2%). PPD subjects met the diagnostic threshold for at least one specific DSM-III-R PD, whereas NPD subjects (a) did not meet the DSM-III-R defined threshold for diagnosis and (b) had fewer than 10 PD features across all disorders. Extensive detail concerning the initial subject selection procedure and sampling is given elsewhere. The 258 subjects consisted of 121 males (47%) and 137 females (53%); 134 (66 female) were in the PPD group and 124 (71 female) were in the NPD group. The mean age of the subjects at entry into the study was 18.89 years (SD = 0.51). All subjects gave voluntary written informed consent and received an honorarium of $50.00 at each wave. Of the initial 258 subjects, 250 completed all three assessment waves and are included in this analysis. Five PPD and three NPD subjects did not complete all three waves. Sample characteristics are summarized in Table 1.

Demographic characteristics of the Longitudinal Study of Personality Disorders sample (N = 250) for subjects available at three waves of the study

PD assessments

As noted above, the LSPD has a prospective multiwave longitudinal design with subjects initially evaluated at three points in time (i.e., first, second, and forth years in college). Although not required for application of individual growth modeling (Singer & Willett, 2003), the LSPD data are balanced, in that all subjects have three waves of data, and are time structured in that everyone was assessed repeatedly on the same three-wave schedule, although the time between assessments varies from case to case. Interview assessments were conducted by experienced PhD or advanced MSW clinicians. Finally, because the LSPD is a naturalistic prospective study, subjects were free to seek psychological treatment of their own accord.

PDE DSM-III-R screen (PDE-S). This is a 250-item self-administered true–false PD screening inventory developed by A. W. Loranger. The diagnostic efficiency and psychometric properties of the IPDE-S in a two-stage screen application were described previously (Lenzenweger et al., 1997).

PDE. The PDE (Loranger, 1988), now known as the IPDE (Loranger, 1999; Loranger et al., 1994), was used in this study. It assessed the DSM-III-R criteria for the Axis II disorders. Clinically experienced interviewers received training in IPDE administration and scoring by Dr. A. W. Loranger and were supervised throughout the project by the first author (M.F.L.), who was blind to the subjects' identity, putative PD status, and all prior assessment information. The interrater reliability for IPDE assessments was excellent at all three waves, ranging between .84 and .92 for all PD dimensions. The interviewers were (a) blind to the putative PD group status of the subjects and (b) blind to all prior LSPD PD assessment data, and (c) the same interviewer did not assess the same subject more than once. The BPD dimensional score from the PDE was used in this analysis.

MCMI—II. The MCMI-II (Choca & Van Denburg, 1997; Millon, 1987) is the well-known 175-item, true–false, self-administered inventory designed to coordinate with DSM-III-R Axis II PDs. The MCMI-II possesses excellent psychometric properties and yields separate dimensional scores for each of the 11 Axis II PDs (Millon, 1987). The dimensional score used in this analysis was the MCMI-II BPD score (base rate [BR] scores; Millon, 1987).

Neurobehavioral indicators assessments.

Neurobehavioral systems indicators for the DLM. The major dimensions hypothesized to be reflective of the neurobehavioral systems that Depue and Lenzenweger (2001, 2005) posit correspond approximately with those discussed by Tellegen (1985) and assessed by his Multidimensional Personality Questionnaire (Tellegen, 1982). The Tellegen measure, however, was not included in the LSPD at the time of initial data collection, but the LSPD did include another personality measure, the NEO Personality Inventory (NEO-PI). Fortunately, Church (1994) conducted an extensive factor analytical study that compared the Tellegen approach with the NEO-PI factors and he provides a quantitative basis for deriving the Tellegen factors from the NEO-PI item pool. Thus, using the Church (1994) data, one can derive four of the five DLM dimensions from the NEO-PI: agentic positive emotion (agency, incentive motivation), communal positive emotion (affiliation), constraint (nonaffective constraint, neural constraint), and anxiety (negative emotion). The NEO-PI (Costa & McCrae, 1985) is the original version of the well-known, Likert-style self-report measure of the personality traits that are part of what has been termed the five factor model. The NEO-PI scores for each of the major personality dimensions have been shown to be internally consistent and reliable over time; the instrument has generally excellent psychometric properties (Costa & McCrae, 1985). Thus, using the approach of Church (1994), approximations of those dimensions hypothesized in the DLM could be extracted from the NEO-PI; we note these dimensions are not isomorphic with the dimensions in the five factor approach.

The fear component of the DLM was extracted from a different self-report inventory known as the Emotionality, Activity, and Sociability (EAS) Adult Temperament Scale. The EAS Adult Temperament Scale (Buss & Plomin, 1984), which is for use with adults, is a 20-item self-report measure of the three major temperament constructs: emotionality (E; which breaks down further to “fear” and “distress” subdimensions), activity (A), and sociability (S), which are thought to underlie personality processes in children and adults. For the purposes of this study, we used the fear dimension from the EAS to represent the fear component of the DLM.

Axis I disorders assessments.

Structured clinical interview for DSM-III-R—Nonpatient version (SCID-NP). This is the well-known semistructured DSM-III-R Axis I clinical interview for use with nonpatients (Spitzer, Williams, Gibbon, & First, 1990). The SCID interview was performed first and followed by the PDE.

Statistical analysis

We used IGC analysis to investigate change in BPD features over time. This method of analyzing within-subject change was popularized by Rogosa and colleagues (Rogosa et al., 1982; Rogosa & Willett, 1985), and represents the most powerful way to assess change in a continuous dimension over time within subjects (Bryk & Raudenbush, 1987; Raudenbush & Bryk, 2002; Rogosa & Willett, 1985; Singer & Willett, 1988, 2003; Willett et al., 1998). The dependent variables used in these analyses was the number of BPD features rated as present on the PDE or the BPD score on the MCMI-II, both of which represent continuous dimensional scores for the disorder. Dimensional measures of BPD ensured the greatest sensitivity to the investigation of stability and change (qualitative diagnoses would not be appropriate for a study of change in this framework).

The IGC approach remains underutilized in developmental psychopathology (especially in the study of PDs); therefore, we describe the method here in some detail. The IGC method hypothesizes that, for each individual, the continuous outcome variable (e.g., the number of BPD features) is a specified function of time, called the individual growth trajectory, plus error. This trajectory is often specified as a simple linear function of time, in which case it contains two important unknown individual growth parameters (an intercept and a slope) that determine the shape of individual true growth over time. The individual intercept parameter represents the net “elevation” of the trajectory over time; that is, the true mean level of the PD features for the individual at the onset of the study (or, alternatively, whenever the origin of the time scale has been defined). The individual slope parameter represents the rate of change over time and here it is the within-person rate of change in BPD features over time. Despite the obvious theoretical importance of individual slope in studies of change, it has not been widely examined in research on psychopathology. Once an individual growth trajectory has been specified (at Level 1) to represent the individual change over time, a Level 2 model can be specified to describe the investigators' hypotheses about the way that the individual growth parameters contained in the Level 1 model are related to between-subjects factors (e.g., subject gender, diagnostic group, neurobehavioral dimensions).

The IGC approach has several advantages. First, interindividual variability in assessment intervals can be tolerated in specifying the individual trajectories, unlike repeated-measures MANOVA. Second, although estimation of an individual growth trajectory and its precision generally requires more than two observations on the individual over time, the IGC modeling approach is highly tolerant of missing data, permitting subjects who have incomplete data to participate in the analyses. Third, using modern software, the IGC modeling approach simultaneously uses data on all individuals at every time point to concurrently investigate both within- and between-individual change, with concomitant improvements in precision and power. Fourth, when lengthy multiwave data are available, IGC modeling permits the flexible specification and rich investigation of nonlinear individual change over time.

In our analyses, the hypothesized Level 1 and Level 2 statistical models were fitted simultaneously to the LSPD data using full maximum likelihood estimation and the HLM-6 computer program. We conducted our analyses sequentially. First, we conducted a set of “unconditional” growth analyses (Singer & Willett, 2003) in which we posited a linear individual change trajectory at Level 1, but did not attempt to predict interindividual variation in the growth parameters by between-subject factors. Such unconditional analyses are useful for partitioning the outcome variation into variance components that describe the net variation in slope and intercept across individuals. Second, a set of “conditional” analyses were conducted in which we examined systematic interindividual differences in intercept and slope as a function of a set of between-subject predictors, namely, study group (PPD vs. NC), subject gender, and age at entry into the study (years) as well as the neurobehavioral systems indicator dimensions noted above. These predictors yield “fixed effects” in the prediction of the slope and intercept values retained from the Level 1 analysis. The fitting of the Level 2 model also yields estimates of residual variance that describe remaining interindividual variability in the individual slopes and intercepts (as well as their covariance) after accounting for the hypothesized fixed effects, giving rise to the variance components (i.e., σ₀², σ₁², σ₀₁; see Appendix A for technical detail).

Improvement in model fit was assessed by comparison of deviance statistics. Fixed effects and variance components were tested for statistical significance using the provided z statistics (two tailed). Effect size estimates for the fixed effects are represented by the effect size r (Rosenthal & Rosnow, 1991).

Clinical characteristics of the sample

As reported previously (Lenzenweger, 1999; Lenzenweger et al., 2004), the lifetime DSM-III-R Axis I diagnoses (Table 2) of the study subjects are for definite and probable disorders. Eighty-one (62.8%) of the PPD subjects received an Axis I diagnosis compared to 32 (26.4%) NPD subjects (χ² = 33.30, df = 1, p < .00001). Forty-one (31.8%) PPD subjects versus 21 (17.4%) NPD subjects reported a prior history of treatment by Wave III (χ² = 6.97, df = 1, p < .008). Finally, as of Wave III, 16% of the sample received a probable or definite diagnosis for at least one Axis II PD (or PD not otherwise specified). We note this figure is higher than that initially reported by Lenzenweger et al. (1997), which was 11%. This is so because some additional subjects developed a PD during the study period beyond those who were diagnosable at Wave I.

Lifetime DSM-III-R Axis I SCID-NP diagnoses (definite and probable) for a sample of 250 subjects

Assessment schedule characteristics

The PD features of each of the 250 study subjects were assessed three times over the 4-year period. The average age of study subjects at the assessment waves were 18.78 years (SD = 0.51) for Wave I, 19.83 years (SD = 0.54) for Wave II, and 21.70 years (SD = 0.56) for Wave III. The time between assessments for each subject was calculated in years, using each subject's date of birth and exact assessment date, and then centered on age at entry into the study for each study subject (with age at entry being included as a predictor at Level 2). The mean time between entry into the study (Wave I) and Waves II and III was 0.95 years (SD = 0.14) and 2.82 years (SD = 0.23), respectively. Centering the assessment intervals on age at entry and including age at entry as a predictor at Level 2 accounts for each subject's unique chronological age when he/she began the study and causes the individual Level 1 intercepts to represent the true value of the Wave I assessments as the subjects' “initial status.” Taking age at entry into the study into account is theoretically important because it helps to account for differences in the developmental level of the study subjects.

IGC modeling of BPD features

The IGC modeling analyses were performed separately for the two outcome variables: PDE BPD features and MCMI-II BPD dimensional score. The heterogeneity in the individual growth trajectories for the PDE and MCMI-II assessments is impressive and they are plotted, using an exploratory ordinary least squares approach, for BPD features in Figure 1.

Ordinary least squares individual growth trajectories for BPD features in 250 subjects over the study period. The time is reported in years since the beginning of the study, and it is centered for each subject using the subject's age at entry into the study. (a) The PDE assessments and (b) the MCMI-II assessments.

Unconditional analyses: Is there change present in the data for BPD?

First, an unconditional growth model (i.e., containing no Level 2 predictors) was fitted for both BPD dimensions, providing estimates of the average elevation and rate of change parameters, and their natural variation, across all subjects on entry into the study. The results of these fits are contained in Table 3, which contains estimates of the fixed effects and variance components for the unconditional growth trajectories for each of the two BPD dimensions. The estimated average elevation of the BPD individual growth trajectories on entry into the study (intercepts) differed significantly from zero for both BPD dimensions (p < .001; representing “large” effects). In addition, each intercept for the PDE and MCMI-II analyses contained significant variability (σ₀²; p < .001) available for prediction at Level 2 in subsequent conditional models. The goodness of fit (deviance) statistic, which indexes model fit, is also reported for both the PDE and MCMI-II data for the unconditional models.

Parameters of the unconditional growth (baseline) model of borderline personality disorder feature change from a prospective multiwave longitudinal perspective (N = 250)

The estimated average rates of change also differed significantly from zero for both of the BPD dimensions, indicating that considerable change over time was evident in the BPD data (both the PDE and MCMI-II revealed “medium” effects for slope). In addition, the variance components associated with rate of change (σ₁²), were statistically significant suggesting that there were significant amounts of variation in change that could potentially be predicted in subsequent Level 2 models. Finally, the estimated slopes from the unconditional growth analyses suggest insights with respect to the rate at which BPD features actually change with time. Specifically, we estimated that total BPD features decrease by 0.14 BPD features on the PDE dimensional score and 1.90 BPD units on the MCMI-II dimensional score with each passing year, both representing medium-sized effects.

Conditional analyses: Distinguishing the individual growth trajectories by subject's group, gender, age at study entry, and neurobehavioral systems indicators

Next, in conditional analyses, we introduced Level 2 predictors to explain any between-person variation in the individual elevation and rate of change parameters. The primary between-subjects Level 2 factors of interest were the group (PPD vs. NPD) membership, subject gender, and baseline values on each of the personality dimensions reflective of the primary DLM components. In addition, we included each subject's age at entry into the study as a predictor at Level 2 to account for interindividual variation in change associated with actual age (i.e., developmental level). For both the PDE and MCMI-II models fitted, intermodel comparisons of goodness of fit (deviance) statistics revealed that the inclusion of group, gender, age at study entry, and the personality dimensions as Level 2 predictors significantly improved the fit beyond that achieved in the unconditional growth models (both p < .001). The results of the conditional analyses are presented for the PDE and MCMI-II in Tables 4 and 5, respectively. These tables include estimates of the fixed effects and variance components associated with each Level 2 predictor (study group, gender, age at entry), agentic positive emotion (agency, incentive motivation), communal positive emotion (affiliation), constraint (nonaffective constraint, neural constraint), anxiety (negative emotion), and fear, the approximate p value for testing that these effects are zero in the population as well as an estimate of the effect size r for the effects, and a deviance statistic (−2 log likelihood) for the model. Tables 4 and 5 also contain estimates of the variance components from the Level 2 models, which were also tested for statistical significance.

Predicting interindividual differences in change of borderline personality disorder in multiwave perspective PDE (N = 250)

Predicting interindividual differences in change of borderline personality disorder in multiwave perspective MCMI-II (N = 250)

For the PDE BPD dimension, with respect to elevation of the individual growth trajectories, fear, anxiety (negative emotion), nonaffective constraint, and agentic positive emotion (agency, incentive motivation) were statistically significant predictors of individual elevation parameters in BPD features (all p < .01). Of these Level 2 predictors of elevation, anxiety (negative emotion), and constraint were associated with the largest effect sizes. Higher levels of negative emotion and agentic positive emotion were associated with higher BPD elevations, whereas lower levels of constraint were associated with higher BPD elevations. Higher levels of fear were associated with lower BPD elevations. The variance component estimate for elevation (σ₀²) indicated that there remained significant variation in elevation that could be modeled beyond the predictors that we had selected.

For the MCMI-II BPD dimension, with respect to elevation of the individual growth trajectories, group membership (PPD vs. NC), age at entry, anxiety (negative emotion), nonaffective constraint, and agentic positive emotion (agency, incentive motivation) were all statistically significant predictors of individual elevation parameters in BPD features (all p < .02 or less). Of these Level 2 predictors of elevation, group membership, anxiety (negative emotion), constraint, and agentic positive emotion medium to large largest effect sizes. Group membership (i.e., PPD status) was associated with higher BPD elevations, whereas younger age at entry was associated with increased BPD elevations. With respect to the personality dimensions, higher levels of negative emotion were associated with higher BPD elevations, lower levels of constraint were associated with higher BPD elevations, and lower levels of agentic positive emotion were associated with higher BPD elevations (all representing “medium” effect sizes). The variance component estimate for elevation (σ₀²) indicated that there remained significant variation in elevation that could be modeled beyond the predictors that we had selected.

The critical growth parameter for investigating stability and change in BPD features over time is the individual slope parameter as it directly indexes the rate and direction of individual change over time. In the Level 2 prediction of slope for the PDE data, fear, anxiety (negative emotion), constraint, and agentic positive emotions were significantly predictive of the rate of change in BPD features (all p < .05 or less). The effects were such that higher baseline levels of anxiety (negative emotion) and agentic positive emotion were predictive of declines in BPD features on the PDE over time. It is interesting that higher baseline levels of fear and constraint actually predicted small effect increases (or, in some cases, less steep declines) in BPD features over time on the PDE. With the exception of negative emotion (medium effect), each of these predictors represented a small effect in the prediction of change. The variance component estimates for rate of change (σ₁²) indicated there was also additional significant variation in elevation that could still be modeled beyond the predictors that we had selected.

Considering the MCMI-II data, in the Level 2 prediction of slope, group status (i.e., PPD classification) and agentic positive emotion were significantly predictive of change in BPD scores over time. Anxiety (negative emotion) was associated with change in BPD features, but at the trend level (p < .07). The effects were such that being a PPD classified subject was associated with a decline in BPD scores and higher levels of agentic positive emotion predicted declines in BPD scores; higher anxiety (negative emotion) levels were predictive (trend level) of declines in BPD scores. The variance component estimates for rate of change (σ₁²) indicated there was also additional significant variation in elevation that could still be modeled beyond the predictors that we had selected.

Finally, in sensitivity analyses suggested by the arithmetic and distributional properties of the dependent variable (as a count of features), we refitted all of our unconditional and conditional models replacing the existing outcome by its square root. This yielded a pattern of results completely consistent with those reported here for the untransformed PD variables.