Comparison of alternative models for personality disorders, II: 6-, 8-, and 10-year follow-up

Abstract

Background

A number of conceptual models have been considered for the assessment of personality pathology in DSM-5. This study sought to extend the findings of Morey et al. (2007) to compare the long-term predictive validity of three such models: the Five-Factor Model (FFM), the SNAP, and DSM-IV personality disorders.

Method

An inception cohort from the Collaborative Longitudinal Personality Disorder Study was followed for 10-years. Baseline data were used to predict long-term outcomes, including functioning, Axis I psychopathology, and medication use.

Results

Each model was significantly valid, predicting a host of important clinical outcomes. Lower-order elements of the FFM system were not more valid than higher order factors, and DSM-IV diagnostic categories were less valid than dimensional symptom counts. Overall, approaches that integrate normative traits and personality pathology proved to be most predictive, as the SNAP, a system that integrates normal and pathological traits, generally showed the largest validity coefficients overall, and the DSM-IV personality disorder syndromes and FFM traits tended to provide substantial incremental information relative to one another.

Conclusion

DSM-5 personality disorder assessment should involve an integration of personality traits with characteristic features of personality disorders.

The DSM-IV-TR (APA, 2000) conceptualization of personality disorders as a group of categorical entities has led to an appreciable research base on a number of these constructs (e.g, Blashfield Intoccia, 2000), pointing to their clinical utility and predictive validity. Nonetheless, there is broad dissatisfaction with this representation of personality disorders (Krueger et al., 2007; Widiger et al., 2009), and various studies have indicated that dimensional alternatives may be more reliable (Heumann Morey, 1990; Widiger Coker, 2002) and valid (Morey et al., 2007; Markon et al., 2011) than DSM-IV categories. However, agreement on the selection of a particular dimensional model to replace the DSM-IV concepts remains limited. Among the alternative proposals are a) modifying categorical DSM-IV constructs with dimensional representations (Skodol et al., 2005); b) replacing personality disorder diagnoses with an assessment of normative traits thought to underlie personality disorder symptomatic expression (Widiger Trull, 2007), such as those of the Five-Factor Model (FFM); and c) assessing traits that are thought to span normal and abnormal personality processes, such as those represented on the Schedule for Nonadaptive and Adaptive Personality (SNAP: Clark, 1993) or the Dimensional Assessment of Personality Pathology (DAPP: Livesley et al., 1998) systems.

In an earlier article (Morey et al., 2007), we examined the criterion validity of five different models for representing personality disorder in the Collaborative Longitudinal Personality Disorders Study (CLPS, Gunderson et al., 2000), examining validating variables that included antecedent variables of potential relevance to etiology and pathogenesis, variables assessing baseline functioning, as well as 2- and 4-year outcomes. The results of that investigation yielded two broad conclusions. First, there was consistent evidence that dimensional characterizations of personality disorder demonstrated greater concurrent and predictive validity relative to DSM-IV categorical diagnoses. Second, it appeared that models of personality pathology could be distinguished between those that more heavily represent normative trait dispositions (such as the FFM) and those that focus on more maladaptive behavioral manifestations of personality pathology (such as the DSM-IV concepts), and that the latter tended to demonstrate greater concurrent validity, whereas the former demonstrated superior predictive validity. Such findings have led to a proposal for personality disorders DSM-5 that reflects a “hybrid” combination of personality traits and disorder types (Skodol et al., 2011).

The purpose of the present report is to present findings from this project at six, eight, and ten years of follow-up, to determine whether the trends observed over the first four years of this longitudinal study generalize to more distal outcomes. For example, it is possible that some personality characteristics might be predictive of remission, and thus informative for understanding intermediate outcome, but different features might be related to relapse and consequently predictive of longer-term outcomes. As in the previous paper (Morey et al., 2007), we provide a comprehensive comparison of these personality models with respect to predictive and incremental validity, using validating markers such as functioning, treatment utilization, suicidal behavior, and Axis I psychopathology. We hypothesized that results would extend our previous observation that normative personality traits and maladaptive behavioral features represent related but incompletely overlapping phenomena that are each incrementally important in understanding dysfunction related to personality. Furthermore, we examine implications of our findings for the integration of traits and disorders as formulated for the representation of personality disorders in DSM-5.

Method

Participants were enrolled in the Collaborative Longitudinal Personality Disorders Study (CLPS; Gunderson et al, 2000). The initial 668 participants ranged in age from 18 to 45 at the time they were recruited from one of the four clinical CLPS sites. Patients met criteria at baseline for at least one of four personality disorders (avoidant, borderline, obsessive-compulsive, or schizotypal) or major depression without personality disorder. All were previously in treatment or treatment-seeking. Exclusion criteria were active psychosis, history of psychotic disorder, acute substance abuse or withdrawal, or significant confusion. At baseline, the mean age was 32.7 (S.D. = 8.1), 64% were women, and 75% were self-described as white.

Assessment

Patients were interviewed at baseline by experienced clinicians using the Diagnostic Interview for DSM-IV Personality Disorders (Zanarini et al., 2000). They also completed the self-report NEO-PI-R (Costa McCrae, 1992), a 240-item measure of FFM traits and 30 lower-order facets; and the 375-item Schedule for Nonadaptive and Adaptive Personality (SNAP: Clark, 1993), a measure of 15 normative and pathological personality traits. These instruments were re-administered at 10-year follow-up. Due to attrition, 545 (82%) of the original 668 patients completed interviews at the 6-year follow-up, with 479 (72%) followed and 8 years and 431 (65%) interviewed at 10-year follow-up.

Several instruments administered at the 6-, 8-, and 10-year follow-up assessed functional outcomes. Interviews included the Global Assessment of Functioning Scale (GAFS); variables from the Longitudinal Interval Follow-up Evaluation (Keller et al, 1987) that assess social, occupational, and recreational functioning; follow-up assessments of Axis I diagnoses; and assessments of the occurrence of hospitalizations, medication use, and suicidal behaviors from study year 4 to study year 10. These variables were available for the participants who completed the follow-up assessments. A self-report measure of depression, the PAI-Depression (Morey, 1991) was also added at study years 6, 8 and 10, as was a self-report measure of interpersonal problems, the Inventory of Interpersonal Problems (IIP: Alden, Wiggins, Pincus, 1990), thereby extending the range of assessed outcomes beyond our earlier follow-up .

Analyses

The analytic strategy was consistent with the earlier report (Morey et al., 2007) in comparing five diagnostic models for personality disorder: 1) the ten DSM-IV personality disorders represented as categorical diagnoses (10 variables, coded present/absent); 2) the ten dimensional DSM-IV personality disorders expressed as criteria counts (10 dimensionalized criterion count variables); 3) the 15 SNAP traits; 4) the five FFM higher-order factors, and 5) the 30 FFM lower-order facets. The analyses focused upon three steps. First, replicating the strategy reported in Morey et al. (2007), all variables in each assessment model (i.e., 10 for the DSM models, 15 for the SNAP, and 5/30 for the FFM domains/facets) were simultaneously entered in multiple regression models to predict 6-, 8-, and 10-year outcomes. In addition, the abilities of each model to predict the variables in the other models over time were determined to provide details about how the models relate to one another in ways that might help explain overlapping and non-overlapping contributions between the models with respect to the prediction of outcome variables. Finally, the incremental validity of each model, controlling for each other model, was calculated to further clarify the unique contribution of different personality models to the prediction of outcomes.

As observed in the earlier report, each model has a different number of predictors and it is well-known that, all things equal, models with more predictors generally result in larger coefficients of determination. Thus, as with our earlier investigation, we employed the Predicted Residual Sums of Squares (PRESS) cross-validation method to correct for potential model over-fitting (Stevens, 2002). A PRESS analysis builds a model with data from every participant except the person whose score is being predicted, and this recurs for every participant. Model effect size is estimated based on the observed residuals in the entire sample, providing a cross-validation uninfluenced by over-fitting resulting from larger numbers of predictors in the regression model. Such an approach is particularly useful to control for overfitting artifacts arising from the use of correlated predictor variables (Weisberg, 1985), which is typically the case when studying mental health variables as predictors.

In addition, PRESS-based residualized models were constructed to examine incremental , validity of the three broad models. For example, to determine the incremental validity of the FFM over the DSM criteria counts in predicting GAFS scores, we saved the residual from the DSM PRESS regression models to determine whether the FFM variables could account for any remaining variance in the GAFS after all variance associated with the DSM model had been removed. Similar analyses were conducted using the DSM variables controlling for the FFM domains, and so forth, for all paired comparisons of the three models .

Results

Table 1 shows the uncorrected and cross-validated model effect sizes for predicting outcome variables at 6-, 8-, and 10-year follow-up. The results of these analyses are reported as both multiple correlations and the square root of PRESS r² values. A type I error rate of .001 was selected to adjust for the probability of spuriously significant values given the considerable number of (multiple) analyses. With respect to predictive validity, results were generally consistent with those from the first four years of follow-up as noted in Morey et al (2007). First, there was substantial support for the validity of all five models in predicting a wide range of outcomes, with all models demonstrating significant predictive validity for the majority of outcome variables, and the SNAP model significantly predictive of all outcomes, including those newly assessed, at all three times points. Second, as was the case in Morey et al. (2007), the categorical DSM-IV diagnoses demonstrated appreciably inferior predictive validity relative to their dimensional counterparts, with the dimensional version demonstrating larger PRESS R than categorical diagnosis in 22 of 24 outcomes (sign test, p < .001). Furthermore, and also consistent with Morey et al. (2007), the 5 higher order FFM domains demonstrated better validity upon cross-validation than the 30 lower-order FFM facets (larger PRESS R in 17 of 24 comparisons, sign test p <.04), also replicating previous findings. The relative predictive validity advantage of the SNAP dimensions noted in the Morey et al. (2007) report over the first four years of follow-up, relative to the FFM domains or the DSM dimensional model, was still evident, with the SNAP demonstrating the largest PRESS R of these three models in 14 of 24 outcome variables (binomial p <.01).

Table 1.

Multiple Rs (and PRESS Multiple Rs) for personality models predicting clinical outcomes.

	DSM categories	DSM dimensions	SNAP	FFM domains	FFM facets
6 year
GAF	.41 (.36)	.46 (.43)	.47 (.42)	.36 (.34)	.44 (.32)
LIFE social functioning	.31 (.23)	.37 (.31)	.37 (.30)	.32 (.29)	.37 (.22)
LIFE work functioning	.29 (.22)	.32 (.26)	.31 (.22)	.25 (.21)	.34 (.18)
LIFE recreational functioning	.25 (.16)	.32 (.26)	.37 (.30)	.33 (.29)	.38 (.23)
Number of Axis I disorders	.25 (.18)	.30 (.24)	.32 (.25)	.26 (.22)	.31 (.17)
Number of current meds	.16 (.03)	.15 (.01)	.26 (.16)	.14 (.05)	.39 (.25)
PAI depression	.34 (.27)	.38 (.32)	.51 (.44)	.43 (.398)	.53 (.400)
Mean/Median PRESS R	.21/.22	.26/.26	.30/.30	.26/.29	.25/.23
8 year
GAF	.40 (.35)	.48 (.44)	.54 (.49)	.41 (.381)	.49 (.380)
LIFE social functioning	.33 (.25)	.41 (.36)	.40 (.33)	.31 (.27)	.40 (.24)
LIFE work functioning	.20 (.07)	.27 (.18)	.30 (.19)	.21 (.15)	.32 (.10)
LIFE recreational functioning	.26 (.17)	.33 (.26)	.37 (.29)	.34 (.31)	.40 (.25)
Number of Axis I disorders	.22 (.13)	.26 (.19)	.31 (.24)	.23 (−.18)	.32 (.19)
Number of current meds	.20 (.03)	.24 (.10)	.33 (.19)	.07 (.00)	.42 (.21)
PAI depression	.25 (.04)	.33 (.17)	.52 (.38)	.45 (.39)	.56 (.32)
Mean/Median PRESS R	.15/.13	.24/.19	.30/.29	.19/.27	.24/.24
10 year
GAF	.37 (.31)	.45 (.40)	.52 (.46)	.37 (.34)	.49 (.36)
LIFE social functioning	.28 (.17)	.37 (.30)	.36 (.26)	.28 (.24)	.40 (.23)
LIFE work functioning	.17 (−.15)	.30 (.11)	.32 (.06)	.23 (.07)	.45 (.03)
LIFE recreational functioning	.18 (.03)	.24 (.13)	.43 (.36)	.33 (.29)	.43 (.27)
Number of Axis I disorders	.16 (.01)	.23 (.14)	.26 (.15)	.20 (.15)	.27 (.07)
Number of current meds	.18 (−.06)	.20 (.00)	.30 (.13)	.11 (−.06)	.44 (.21)
PAI depression	.32 (.22)	.38 (.30)	.53 (.46)	.45 (.41)	.53 (.36)
IIP interpersonal problems	.43 (.27)	.41 (.26)	.36 (.35)	.36 (.26)	.52 (.21)
# Suicide attempts	.26 (.13)	.28 (.17)	.23 (.10)	.10 (−.02)	.28 (.07)
# Psychiatric hospitalizations	.30 (.22)	.31 (.24)	.27 (.16)	.12 (.03)	.25 (.00)
Mean/Median PRESS R	.12/.15	.21/.21	.25/.26	.17/.24	.18/.23

Note. Suicide attempts and hospitalizations occurred between years 4 and 10. Significant (p < .001) mutiple correlations are in bold.

Given the relatively limited predictive validity of the categorical DSM-IV and FFM facet models, subsequent analyses focused upon the remaining three models to explore their relationship to one another and the unique contribution of each model independent of the others. Table 2 depicts the ability of the baseline dimensional DSM-IV diagnoses, SNAP traits, and FFM domains to predict the variables in the other models at year 10 follow-up. All models demonstrated significance in predicting each other, which supports the contention that they are accounting for some overlapping variance. The baseline SNAP traits tended to be better at predicting 10-year DSM dimensions than the FFM domains, and better at predicting the 10-year FFM domains than the DSM dimensions; the baseline FFM domains tended to predict the 10-year SNAP traits better than the DSM dimensions.

Table 2.

10-year cross-model predictions in Multiple R and (PRESS Multiple R) metrics.

	FFM traits	SNAP Traits
DSM-IV PDs
Paranoid	.39 (.35)	.48 (.41)
Schizoid	.31 (.27)	.39 (.29)
Schizotypal	.31 (.23)	.54 (.42)
Borderline	.28 (.23)	.39 (.29)
Histrionic	.25 (.19)	.32 (.16)
Narcissistic	.24 (.15)	.25 (.09)
Antisocial	.27 (.21)	.45 (.36)
Avoidant	.44 (.41)	.48 (.42)
Dependent	.26 (.20)	.34 (.20)
Obsessive-Compulsive	.24 (.17)	.32 (.18)
Mean/Median PRESS	.24/.22	.28/.29
	DSM-IV PDs	SNAP traits
FFM Traits
Neuroticism	.42 (.34)	.60 (.54)
Extraversion	.47 (.39)	.64 (.58)
Openness to Experience	.06 (−.06)	.44 (.32)
Agreeableness	.45 (.38)	.56 (.48)
Conscientiousness	.33 (.19)	.57 (.50)
Mean/Median PRESS	.25/.34	.48/.50
	DSM-IV PDs	FFM Traits
SNAP Traits
Negative Temperament	.32 (.21)	.51 (.48)
Mistrust	.48 (.42)	.51 (.48)
Manipulativeness	.37 (.24)	.43 (.40)
Aggression	.40 (.31)	.49 (.46)
Self-harm	.47 (.40)	.45 (.41)
Eccentric Perceptions	.44 (.36)	.18 (.00)
Dependency	.29 (.12)	.34 (.28)
Positive Temperament	.38 (.28)	.57 (.54)
Exhibitionism	.38 (.29)	.48 (.45)
Entitlement	.44 (.36)	.40 (.35)
Detachment	.44 (.37)	.57 (.55)
Disinhibition	.37 (.26)	.50 (.48)
Impulsivity	.26 (.11)	.49 (.46)
Propriety	.28 (.08)	.37 (.32)
Workaholism	.30 (.16)	.38 (.34)
Mean/Median PRESS	.26/.28	.44/.45

Note. Predictor set in columns, predicted dimension in rows. Significant (p < .001) multiple correlations are in bold.

The next set of analyses examined the incremental validity of each model in terms of its unique contribution to prediction beyond that provided by alternative models. In these analyses, PRESS predicted scores were computed for participants for each of the three models, predicting each validity criterion, thus allowing each model to be represented by one independent variable which reflects its estimate of each participant’s criterion score. Then, the part-correlation between these estimates (controlling for the estimates derived from the alternative models) were computed for the various outcome measures. This part-correlation indicates the incremental validity added by the first model in estimating the specified criterion over and above the estimate provided by the second model. Thus, for example, a significant part-correlation between the DSM and 10-year GAF, controlling for SNAP, would indicate that the DSM model provides information about 10 year functional outcome above and beyond that provided by the SNAP traits.

The results of these analyses are reported in Table 3. Statistically significant part-correlations were observed between the SNAP and most of the criteria after controlling for the contribution of the DSM-IV disorder dimensions. However, fewer than half of the part-correlations of the DSM-IV with outcomes after controlling the SNAP are statistically significant. The results also indicated that the SNAP incremented the information provided by the FFM to a greater extent than was true of the converse. It appears as though the DSM-IV model tended to add significant incremental validity to the SNAP in the prediction of GAFS and social functioning, while the FFM did not demonstrate any significant incremental validity over the SNAP dimensions on any outcome variables. Both the FFM and the DSM demonstrated incremental contributions over the other model, in that the FFM incremented the DSM and the DSM also incremented the FFM.

Table 3.

Part correlations of PRESS-derived predicted scores with outcomes.

Tested Model	DSM		SNAP		FFM
Controlled Model	SNAP	FFM	DSM	FFM	DSM	SNAP
6 year
GAF	.21	.32	.23	.30	.17	.07
LIFE social functioning	.19	.21	.19	.17	.18	.14
LIFE work functioning	.17	.19	.11	.12	.12	.08
LIFE recreational functioning	.11	.14	.21	.15	.20	.11
Number of Axis I disorders	.10	.15	.15	.15	.13	.07
Number of current meds	−.04	.01	.16	.15	.05	−.01
PAI depression	.06	.12	.34	.25	.28	.11
Mean/Median R	.11/.11	.16/.15	.20/.19	.18/.15	.16/.17	.08/.08
8 year
GAF	.17	.32	.35	.38	.23	.03
LIFE social functioning	.22	.27	.21	.24	.15	.10
LIFE work functioning	.13	.15	.15	.13	.11	.05
LIFE recreational functioning	.11	.14	.20	.14	.22	.13
Number of Axis I disorders	.08	.13	.18	.17	.13	.05
Number of current meds	.06	.09	.16	.17	−.18	−.17
PAI depression	−.03	−.01	.36	.17	.36	.17
Mean/Median R	.11/.11	.16/.14	.23/.20	.20/.17	.15/.15	.05/.05
10 year
GAF	.16	.30	.35	.39	.21	.10
LIFE social functioning	.20	.23	.18	.19	.16	.12
LIFE work functioning	.10	.09	.04	.03	.05	.06
LIFE recreational functioning	−.03	.00	.34	.25	.25	.08
Number of Axis I disorders	.06	.09	.12	.10	.11	.05
Number of current meds	−.06	.00	.15	.16	−.07	−.10
PAI depression	.07	.11	.36	.26	.29	.12
IIP interpersonal problems	.11	.17	.26	.25	.17	.06
Suicide attempts	.15	.17	.03	.10	−.06	−.06
Psychiatric hospitalizations	.19	.23	.02	.14	−.06	−.04
Mean/Median R	.10/.11	.14/.14	.19/.17	.19/.18	.11/.14	.04/.06

Note. Only the FFM domains and DSM dimensions were considered. Significant (p < .001) incremental correlations are in bold.

Given the apparently overlap between SNAP and FFM models, additional analyses were conducted to explore the extent to which specific DSM-IV personality disorders incremented the predictive validity of an aggregation of these trait-based models. The first step in this process involved the creation of a composite dimensional trait model using a conjoint exploratory factor analysis of study baseline scores on the 15 SNAP dimensions and the 5 FFM domains, using a Principal Axis Factoring method followed by an oblimin rotation. This analysis resulted in five factors with eigenvalues greater than one, cumulatively accounting for 60% of the variance. The structure matrix for this solution is found in Table 4, which indicates that these five factors resemble the broad domains thought to describe the pathological ranges of the FFM (e.g., Widiger, 2011). Then, hierarchical regression analyses were conducted to predict GAF scores at two-year intervals between study baseline and 10-year follow-up; these analyses entered the five factor scores from the conjoint factor analysis as the first block, and then the dimensional symptom counts from the DSM-IV disorders as the second block.

Table 4.

Conjoint Factor Analysis of Baseline SNAP and NEO-PI-R Variables.

	Factor
	1	2	3	4	5
NEO-PI-R Scales
Neuroticism	0.85	−0.36	−0.18	−0.18	0.10
Extraversion	−0.29	0.89	−0.03	0.09	0.21
Openness	−0.11	0.36	−0.15	0.17	0.24
Agreeableness	−0.08	0.10	0.13	0.82	−0.13
Conscientiousness	−0.53	0.26	0.70	0.11	0.29
SNAP Scales
Impulsivity	0.36	0.07	−0.73	−0.24	0.08
Propriety	0.16	−0.05	0.57	−0.19	0.26
Workaholism	0.00	0.07	0.38	−0.09	0.63
Manipulativeness	0.27	0.02	−0.48	−0.65	0.19
Mistrust	0.44	−0.33	0.08	−0.60	0.42
Eccentric Perceptions	0.30	−0.05	−0.08	−0.39	0.56
Aggression	0.38	−0.12	−0.10	−0.68	0.19
Self-harm	0.68	−0.42	−0.20	−0.22	0.20
Detachment	0.24	−0.77	0.03	−0.29	0.18
Exhibitionism	−0.15	0.70	−0.15	−0.22	0.20
Entitlement	−0.22	0.42	0.09	−0.46	0.33
Dependency	0.58	−0.06	−0.16	−0.03	−0.06
Positive Temperament	−0.33	0.72	0.15	−0.05	0.54
Negative Temperament	0.76	−0.28	0.03	−0.33	0.31
Disinhibition	0.30	0.06	−0.83	−0.52	0.07
Eigenvalue	5.10	3.40	2.70	1.70	1.11
Percent Variance	25.50%	16.90%	13.52%	8.53%	5.56%

Note: Loadings above .40 denoted in bold.

The results of these incremental analyses are presented in Table 5. The first row of this Table reveals that the five conjoint SNAP/FFM factors measured at baseline demonstrated multiple correlations ranging from .42 to .51 in predicting GAF scores over the 10 years of the study. Nonetheless, symptom counts from three DSM-IV disorders—Borderline, Schizotypal, and Antisocial--demonstrated significant beta coefficients (i.e., incrementing variance explained by the five conjoint factors) in predicting GAF at every observation point, while a fourth disorder (Schizoid) providing incremental prediction at four of the six intervals. Two disorders, Histrionic and Narcissistic, failed to increment prediction at any of the observed study intervals.

Table 5.

Multiple R for conjoint NEO/SNAP factors and Beta coefficients for PDs predicting GAF.

Predictor	Baseline	2 year	4 year	6 year	8 year	10 year
Factors	.43**	.42**	.42**	.43**	.51**	.46**
Paranoid	−.16**	−.16**	−.08	−.01	.03	.04
Schizotypal	−.27**	−.23**	−.17**	−.22**	−.14*	−.14*
Schizoid	−.16**	−.14**	−.08	−.11*	−.10	−.12*
Antisocial	−.24**	−.24**	−.14*	−.16*	−.17**	−.22**
Borderline	−.34**	−.27**	−.12*	−.13*	−.10*	−.12*
Histrionic	−.06	−.07	.08	.04	.02	−.02
Narcissistic	−.05	−.03	.01	.01	.07	−.02
Avoidant	.02	.01	.10*	.03	.05	.07
Dependent	−.09*	−.08*	.07	.02	.04	.00
OC	.13*	.10*	−.01	.05	.06	.04

^*p < .05

^**p < .001

Discussion

One of the central aims of the CLPS study was to compare the validity of three general models of personality pathology (and variants of two of these models) in predicting outcome in personality disorder over a 10-year period. With respect to this aim, it is important to note that all models (categorical and dimensional DSM-IV, SNAP, and factor and facet level FFM) demonstrated significant predictive validity with respect to both intermediate (as shown in Morey et al., 2007) and longer-term outcome, as shown here. However, although each model demonstrated predictive validity, it was also the case that all of the dimensional models examined consistently outperformed the DSM-IV categorical personality disorder diagnoses, including a dimensional representation of the DSM-IV disorders. Thus, the data presented here and in the earlier report (Morey et al., 2007) provide consistent evidence that any of the examined dimensional models would provide greater predictive utility than the DSM-IV categories, supporting the move toward a dimensional system for classifying personality disorders in DSM-5.

Although the evidence seems clear that the categorical representation of personality disorder presented in DSM-IV leads to significant loss of predictive information relative to a dimensional representation of the same information, the data concerning the relative validity of the alternative dimensional models examined in this study are less clear. As was the case in our earlier study (Morey et al., 2007) as well as in a subsequent study utilizing a community sample (Grucza Garza, 2007), it appeared that use of lower order features in a normative trait hierarchy (i.e., the 30 FFM facets) failed to improve upon the higher order factors (i.e., the five FFM domains) in predicting outcome upon cross-validation. While the potential gain in precision afforded by the use of lower-order traits may hold promise, it is apparent that some of this gain may be illusory, with enhanced statistical prediction reflecting an overfitting of a particular data set. Given the increased complexity associated with the use of larger numbers of lower-order traits within a descriptive personality system, there is a need to demonstrate that this complexity is offset by a true increment in validity above that provided by the higher-order trait domains—which did not appear to be the case in this study. As the DSM-5 trait proposal (e.g., Skodol et al., in press) includes lower-order facets as well as higher order domains, it will be important to examine this issue with respect to that proposal.

When comparing the dimensionalized DSM-IV, FFM domains, and SNAP dimensions, it appeared that the SNAP model tended to demonstrate the greatest capacity to predict outcome. The SNAP was also able to significantly increment the other models in predicting a wide range of criteria, while the opposite was not always true. This finding replicates our earlier observations (Morey et al., 2007). Since the SNAP was developed to be a hybrid model of personality disorder composed of both general personality traits (which the SNAP describes as “adaptive” traits) as well as trait features of particular relevance to personality disorder(which the SNAP describes as “maladaptive” traits), this hybrid nature may account for its enhanced predictive capacity. Further supporting this view, , it appears that the FFM and the DSM models increment one another in clinical prediction. However, it should also be noted that the limited increment in predictive validity of the FFM over the SNAP may derive from differences in the instruments used to assess these models, with the SNAP providing greater information in the pathological range of constructs similar to those assessed in the more normative range by the NEO-PI-R. Nonetheless, the comparison of the DSM model with the SNAP and FFM suggests that the DSM concepts increment those models in clinical prediction of areas such as global functioning, social functioning, and future hospitalizations.

One important extension of the 4-year CLPS results (Morey et al., 2007) involved the time course of the predictive capacity of the different models. The pattern over the first four years of the study indicated that the validity of the DSM dimensions was quite high at baseline, but diminished relatively rapidly over the four years, while the FFM traits sustained a lower but more consistent predictive capacity over the same interval. Although this pattern might produce the expectation that the predictive capacity of the DSM disorders would continue to diminish and ultimately asymptote below that of traits, the data from years 6 through 10 indicate otherwise. The validity of the DSM dimensions for predicting long-term functioning remained as high, if not higher, than the FFM traits, suggesting that aspects of the DSM-IV criterion sets are indeed capturing enduring as well as more evanescent problems. Furthermore, the finding of incremental validity of the DSM (in contrast to the FFM) over the SNAP dimensions in the prediction of outcomes such as global functioning, suicidal behavior, and need for hospitalization suggests that those incremental aspects of the DSM concepts provide prediction of variables that are of considerable clinical relevance. This incremental validity supports perspectives of writers such as Shedler et al. (2010) who have argued that a solely trait-based approach to personality disorder diagnosis might be less clinically useful in certain respects than [AS1]DSM-IV concepts.

This possibility raises an important issue for the consideration of personality trait and disorder concepts in DSM-5—if a trait based model is incorporated into the new diagnostic manual, which existing DSM-IV concepts are important to retain to increment the validity of information provided by those traits? The results described in Table 5 provide important data addressing this issue. It is noteworthy that a conjoint factor analysis of the SNAP and NEO-PI-R instruments produced a factor structure similar to the pathological trait model proposed for DSM-5 (i.e., Krueger et al., 2011), and that these conjoint factors provided appreciable prediction of later functioning over intervals as long as ten years. However, three DSM-IV disorders in particular—Borderline, Schizotypal, and Antisocial—each significantly incremented these conjoint factors at every measurement interval. These results suggest that these diagnostic concepts provide valid information above and beyond that provided by the conjoint trait model, supporting the need to retain these concepts even if such a trait model were implemented. It is important to note that, at least for Borderline Personality, similar conclusions have been derived from other data sets (Morey Zanarini, 2001; Distel et al., 2009). On the other hand, certain DSM-IV disorders such as Histrionic Personality Disorder appeared to provide no prediction of functioning beyond that offered by a pathological trait model. It should be recognized that the CLPS study focused upon four disorders, and thus lack of incremental predictive validity could be potentially related to the representation of these disorders in this sample, although it should be noted that Antisocial Personality criteria demonstrated incremental predictive validity even though it was not a focal study group.

The results of this study have important implications for DSM-5, and are consistent with the efforts in DSM-5 to develop a dimensional personality system that integrates personality trait, disorder, and core pathology concepts (Skodol et al., 2011a; Skodol et al., 2011b), although the hybrid dimensional model of personality and personality disorder assessment and diagnosis proposed for DSM-5 does not include the exact measures or models tested in this study. First, all dimensional models in this study measured at baseline outperformed the DSM-IV categorical approach in predicting important clinical outcomes from 6 to 10 years into the future. These results are consistent with our earlier report (Morey et al., 2007) in supporting the shift toward dimensions in DSM-5. Second, personality trait variables predicted outcomes as well as or better than DSM-IV PDs, supporting the inclusion of a pathological trait model in the DSM-5 (e.g., Krueger et al., 2011). Third, because certain DSM-IV PD dimensions incremented such trait dimensions in predicting outcomes out to 10 years of follow-up, the decision to include both disorders and pathological traits in the DSM-5 hybrid model is supported. Each approach contributed unique information about prognosis, which would argue against exclusively trait-based or exclusively disorder-based models. Finally, our findings that lower-order FFM facets did not increase the predictive validity of the higher order factors raise questions about the levels at which personality traits should be assessed in DSM-5, as the proposed model includes both lower order facets as well as broad trait domains (Krueger et al., 2011). Additional research on the reliability and validity of these different hierarchical levels will be particularly important, as will research to determine whether the predictive validity of the systems assessed here can be sustained when implemented in routine clinical practice. It should be noted that there is evidence to suggest that clinicians can provide dimensional trait ratings that provide results similar to those obtained using questionnaires such as the SNAP and NEO-PI-R (Miller et al., 2010). Nonetheless, the results described here document that careful assessments of both traits and disorder criteria provide considerable long-term prediction of clinically useful information such as functional outcome and treatment utilization.