Authors: Richard E. Zinbarg (1Psychology Department, Northwestern University, University of California – Los Angeles.), Susan Mineka (1Psychology Department, Northwestern University, University of California – Los Angeles.), Lyuba Bobova (1Psychology Department, Northwestern University, University of California – Los Angeles.), Michelle G. Craske (2Psychology Department, University of California – Los Angeles.), Suzanne Vrshek-Schallhorn (1Psychology Department, Northwestern University, University of California – Los Angeles.), James W. Griffith (1Psychology Department, Northwestern University, University of California – Los Angeles.), Kate Wolitzky-Taylor (2Psychology Department, University of California – Los Angeles.), Allison M. Waters (2Psychology Department, University of California – Los Angeles.), Jennifer A. Sumner (1Psychology Department, Northwestern University, University of California – Los Angeles.), Deepika Anand (1Psychology Department, Northwestern University, University of California – Los Angeles.)
Categories: Article, neuroticism, cognitive vulnerability, anxiety disorders, mood disorders, substance use disorders
Source: Clinical psychological science : a journal of the Association for Psychological Science
Authors: Richard E. Zinbarg, Susan Mineka, Lyuba Bobova, Michelle G. Craske, Suzanne Vrshek-Schallhorn, James W. Griffith, Kate Wolitzky-Taylor, Allison M. Waters, Jennifer A. Sumner, Deepika Anand
Neuroticism and several other traits have been proposed to confer vulnerability for unipolar mood disorders (UMDs) and anxiety disorders (ADs). However, it is unclear whether the associations of these vulnerabilities with these disorders are attributable to a latent variable common to all vulnerabilities, more narrow latent variables, or both. Additionally, some suggest that neuroticism predicts UMDs, ADs, and substance use disorders (SUDs) with comparable strength whereas others hypothesize that neuroticism is more strongly related to UMDs and ADs. We tested hypotheses about the factor structure of several vulnerabilities and the prospective associations of these latent variables with initial onsets of UMDs, ADs, and SUDs over three years in 547 participants recruited as high school juniors. Although a general neuroticism factor predicted SUDs, it predicted UMDs and ADs more strongly and especially predicted comorbid UMDs and ADs. There was also mixed support for specific associations involving more narrow latent vulnerabilities.
Neuroticism (N) has been proposed as a common vulnerability for ADs and UMDs (e.g., Eysenck, 1967; Gray & McNaughton, 2000). Other personality traits and cognitive style variables have been hypothesized to be more narrow vulnerabilities for either ADs or UMDs. For example, several cognitive style variables have been proposed to be vulnerability factors for UMDs (e.g., Abramson, Metalsky, & Alloy, 1989; Beck, 1967, 1983; Blatt & Zuroff, 1992; Clark & Beck, 1999). These include dysfunctional attitudes (including the need for approval and need for achievement), negative inferential style (the tendency to interpret negative life events as having stable and global causes which lead to negative consequences), sociotropy (having heightened needs for support and acceptance) and autonomy (being excessively concerned with achievement issues and being highly self-critical). In contrast to dysfunctional attitudes, negative inferential style, sociotropy and autonomy, anxiety sensitivity (AS) – the belief that anxiety and physical sensations of anxiety are harmful – has been hypothesized to be a risk factor for ADs in general and panic disorder (PD) in particular (Reiss & McNally, 1985).
It is already known that most of these hypothesized vulnerability factors do prospectively predict depressed mood and initial onsets of major depression (e.g., Alloy et al., 2006; Kendler, Kuhn & Prescott, 2004; Klein, Durbin & Shankman, 2009; Lewinsohn, Joiner & Rohde, 2001). We also know that one or more facets of AS prospectively predict the onset of panic attacks (e.g., Hayward, Killen, Kraemer & Taylor, 2000; Schmidt, Lerew & Jackson, 1997, 1999), worry about panic (Schmidt, 1999), and ADs considered as a group (Schmidt, Zvolensky & Maner, 2006). In addition, it is already established that AS has unique associations with anxiety symptoms above and beyond measures of broader constructs such as N (e.g., Eke & McNally, 1996; Rapee & Medoro, 1994; Schmidt et al., 1999).
There are, however, many unanswered questions about N, dysfunctional attitudes, negative inferential style, sociotropy, autonomy and AS and their prospective associations with psychopathology. For example, does the cognitive vulnerability conferred by elevations on dysfunctional attitudes, negative inferential style, sociotropy or autonomy predict UMDs significantly more strongly than ADs? Similarly, does AS predict ADs significantly more strongly than UMDs? And, does N prospectively predicts initial onsets of ADs other than post-traumatic stress disorder (e.g., Breslau & Schultz, 2013)?^1^
There is also theoretical disagreement regarding the nature of N and its associations with psychopathology. Within Gray’s reinforcement sensitivity theory (RST; e.g., Gray & McNaughton, 2000), N is hypothesized to be specifically associated with sensitivity to cues for punishment, frustrative non-reward and conflict (but not to cues for reward and relieving nonpunishment). As such, RST predicts that N should be more strongly associated with internalizing than with externalizing psychopathology, with the latter involving a stronger contribution from reward circuits (e.g., Gray & McNaughton, 2000; Zinbarg & Yoon, 2008). By contrast, after reviewing evidence of cross-sectional associations with many forms of psychopathology including substance use disorders (SUDs), Claridge and Davis (2001, p. 383) concluded that N “is such a universal accompaniment of abnormal functioning (both psychological and biological) that by itself it has little descriptive or explanatory value.” However, whereas it is clear that N predicts SUDs (e.g., Sher, Grekin & Williams, 2005), an unanswered question is whether N predicts UMDs or ADs more strongly than SUDs (or other forms of externalizing psychopathology).
Cognitive vulnerability theorists have rarely considered the possibility that cognitive risk variables might be facets of N. Thus, a hypothesis that appears to be implicitly incorporated into many of their theories is that cognitive risk variables are either unrelated to N or have prospective effects above and beyond those of N. By contrast, at least some N theorists have explicitly incorporated cognitive constructs into their definitions of N. For example, Lilienfeld, Turner, and Jacob (1993) proposed that AS is a facet of N. Similarly, Costa and McCrae (1992) considered irrational ideas to be a facet of N. In addition, Eysenck and Eysenck (1985) considered low self-esteem to be a facet of N and Scheier, Carver and Bridges (1994) noted that pessimism often has been hypothesized to be a facet of N.
Consistent with the hypotheses that cognitive constructs similar to dysfunctional attitudes, sociotropy, autonomy and negative inferential style are facets of N, these cognitive constructs have often shown strong associations with N (e.g., Bagby et al., 2001; Dunkley, Blankstein, & Flett, 1997). Further, negative inferential style, dysfunctional attitudes, sociotropy or autonomy are also associated with at least certain ADs (e.g., Mineka, Pury & Luten, 1995) and some aspects of AS are associated with depression (e.g., Zinbarg, Brown, Barlow & Rapee, 2001). The non-specificity of these associations and correlations of these vulnerabilities with N suggests that the general N factor, at least in part, accounts for the associations between these cognitive risk variables and psychopathology. Indeed, some cross-sectional evidence failed to show unique associations between ADs and UMDs with negative inferential style, dysfunctional attitudes, sociotropy or autonomy above and beyond N (Zinbarg et al., 2010). Thus, whether dysfunctional attitudes, sociotropy, autonomy and negative inferential style have unique and specific predictive effects above and beyond the general N factor is also an open question.
Other gaps in theoretical understanding in this area stem from the fact that the hierarchical structure of the vulnerability factors that we are focused on is likely quite complex and difficult to fully account for or comprehend. For example, a number of factor-analytic studies have suggested that the structure of AS is hierarchical with three group factors (i.e., factors common to some but not all items) and a general factor (e.g., Stewart, Taylor, & Baker, 1997; Zinbarg, Barlow, & Brown, 1997). Moreover, AS is thought to be embedded within a larger hierarchical structure along with N (Lilienfeld, Turner, and Jacob (1993). Thus, we don’t know whether the unique effects of AS that have been demonstrated should be attributed to one (or more) of the AS group factors, the general AS factor, or to factors at both levels of the AS hierarchy. It is also unclear whether the general factor common to all AS items might, in fact, be the general N factor. Similarly, whether dysfunctional attitudes and negative inferential style share an additional common factor that is not the general N factor is an open question. The practice of defining cognitive risk for depression on the basis of elevations on both dysfunctional attitudes and negative inferential style (e.g., Alloy et al, 2000, 2006) is equivalent to defining risk on the basis of a composite of dysfunctional attitudes and negative inferential style. This practice assumes that dysfunctional attitudes and negative inferential style share a factor and interpreting the results in terms of cognitive risk for depression (rather than N) implies that this common factor is distinguishable from the general N factor. Unfortunately, this assumption has not been previously tested.
Another unresolved theoretical question regarding overlap among the personality and cognitive risk factors included in this study stems from the substantial overlap of Sociotropy and Autonomy scales with the Needing Approval and Needing Achievement subscales of the Dysfunctional Attitudes Scale (e.g., Dunkley, Sanislow, Grilo, & McGlashan, 2004; Zuroff, 1994). Indeed, it has been suggested that Sociotropy and Needing Approval were likely indicators of one construct, while Autonomy and Needing Achievement were likely indicators of a second construct (e.g., Dunkley et al., 1997; Ouimette, Klein, Anderson, Riso, & Lizardi, 1994). However, to our knowledge this hypothesis has not yet been tested.
Gender is another variable related to risk for internalizing disorders with females at greater risk for UMDs (e.g., Nolen-Hoeksema & Hilt, 2009) and many ADs (e.g., Craske, 2003). Females also score higher than males on N (e.g., Costa, Terracciano, & McCrae, 2001) including in the present sample (Zinbarg et al., 2010) – as well as on the cognitive vulnerabilities (e.g., Hankin & Abramson, 2001). There is also some evidence suggesting that gender moderates the associations between neuroticism and the emotional disorders. For example, N has been found to be significantly more strongly related cross-sectionally to major depression in males than females (Fanous, Gardner, Prescott, Cancro, & Kendler, 2002). Similarly, in the present sample, N was found to be significantly more strongly related to past diagnoses of UMDs and MDD in males (Zinbarg et al., 2010). Unfortunately, it is unclear whether gender moderates the prospective association between N and MDD (Kendler et al., 2004). Notably, the interpretability of gender differences on N has been questioned on the grounds that N scales may not be invariant across men and women (e.g., Reise, Smith, & Furr, 2001). Indeed, if the factor structure of our risk measures is not highly similar for men and women, the measures would be tapping different constructs for men and women and therefore tests of gender moderation of vulnerability associations with emotional disorders could not be interpreted in a straightforward manner.
The first aim of the present study was to test several hypotheses regarding the prospective associations between the latent variables tapped by our hypothesized risk measures and initial onsets of ADs, UMDs and SUDs. Based on reinforcement sensitivity theory (e.g., Gray & McNaughton, 2000) and cross-sectional evidence (e.g., Claridge and Davis, 2001), it is hypothesized that N is a common risk factor for ADs and UMDs. As such, N should predict ADs in addition to UMDs and should be an especially strong predictor of comorbid ADs and UMDs. We also pitted against each other two contrasting hypotheses regarding the associations between N and psychopathology. According to RST (e.g., Gray & McNaughton, 2000), N should prospectively predict UMDs and ADs more strongly than SUDs. In contrast to this RST hypothesis, the Claridge and Davis (2001) perspective hypothesizes that N should predict SUDs as strongly as UMDs and ADs. Based on theory (e.g., Abramson, Metalsky, & Alloy, 1989; Beck, 1967) and existing prospective evidence (e.g., Alloy et al., 2006), we hypothesized that dysfunctional attitudes and negative inferential style predict UMDs and do so more strongly than ADs. Based on AS theory (e.g., Reiss & McNally, 1985) and earlier prospective evidence (e.g., Hayward, Killen, Kraemer & Taylor, 2000), we hypothesized that one or more AS factors predict ADs and do so more strongly than UMDs. Finally, based on past cross-sectional (Fanous, Gardner, Prescott, Cancro, & Kendler, 2002) and retrospective evidence (Zinbarg et al., 2010), we hypothesized that N is a stronger predictor of UMDs in males than females.
A second aim was to test hypotheses regarding the overlap among the risk factors included here. Thus, based on the practice of defining cognitive risk on the basis of elevations on both dysfunctional attitudes and negative inferential style (e.g., Alloy et al, 2006), we hypothesized that dysfunctional attitudes and negative inferential style share a common factor beyond the general N factor. Based on existing theory (e.g., Lilienfeld, Turner, and Jacob, 1993) and past factor analyses (e.g., Zinbarg, Barlow, & Brown, 1997), we also hypothesized that the hierarchical structure of N includes an intermediate-breadth AS factor in addition to the general (broad) N factor and three AS group (narrow) factors. Based on past correlational evidence (e.g., Dunkley et al., 1997), we hypothesized that Sociotropy and Needing Approval are indicators of one construct, while Autonomy and Needing Achievement are indicators of a second construct. Finally, we tested the hypothesis suggested by Reise and colleagues (e.g., Reise, Smith, & Furr, 2001) that the factor structure of measures of N and its cognitive facets differs meaningfully between males and females.
Testing these predictions has important implications not only for theory but also for preventive interventions. Different preventive interventions may be called for depending on which hypotheses are supported by the data. For example, if only the general N factor has unique predictive power for both UMDs and ADs, then those at risk might benefit most from broad-based preventive interventions for general emotional regulation. By contrast, if only specific risk factors for different disorders have unique predictive power, then more narrowly targeted preventive intervention strategies for specific risk factors might be most valuable. One example of more narrowly targeted prevention programs are those that target AS to reduce risk for panic disorder (e.g., Gardenswartz & Craske, 2001).
Participants (n=547) were recruited into the Northwestern-UCLA Youth Emotion Project (YEP) study from the eleventh grade of two ethnically and socio-economically highly diverse high one in suburban Chicago and the other in suburban Los Angeles. Given that many UMDs, ADs and SUDs have their first onset during late adolescence (e.g., Kessler, Bergland et al., 2005), and that this age range involves changing life roles, this is a useful age range in which to study the onset and course of UMDs, ADs and SUDs (Prenoveau et al., 2011). Eleventh grade students who provided assent and parental consent completed a screening questionnaire – a 22-item version of the N scale of the revised Eysenck Personality Questionnaire (EPQ-R-N; Eysenck & Eysenck, 1975). Students were categorized as low-, medium-, and high-EPQ-R-N scorers, and, when inviting participants into the longitudinal study, we oversampled those classified as high-scorers and maintained equal proportions of females to males across the three EPQ-R-N categories. There were 627 students who completed the baseline assessment, which included an assessment of lifetime Axis I psychopathology using the Structured Clinical Interview for DSM-IV, non-patient edition (SCID-I/NP; First, Spitzer, Gibbon & Williams, 2002).
Low-, medium-, and high-EPQ-R-N scoring participants represented 18.4%, 23.0%, and 58.6% of the sample respectively. The sample was 68.7% female. Participants identified themselves as 48.6% Caucasian, 15.3% Latino, 12.4% African American, 5.2% “other”, 4.5% Asian, .7% Pacific Islander, and 13.2% as having more than one race or ethnicity. Participants had a mean age of 16.9 years (SD = 0.4) at the time of their first interview.^2^ These participants, or subsets of them, have been used in a number of previous publications that tested different hypotheses than those tested here (i.e., Adam et al., 2010; Adam et al., 2014; Craske et al., 2009; Craske et al., 2012; DeSantis et al., 2007; Hauner et al., 2008; Griffith et al., 2010; Griffith et al., 2009; Lewis et al., 2010; Mor et al., 2010; Prenoveau et al., 2009; Prenoveau et al., 2010; Prenoveau et al., 2011; Sumner et al., 2011; Sumner, Mineka, Zinbarg et al., 2014; Sumner, Mineka, Adam et al., 2014; Sumner, Vrshek-Schallhorn et al., 2014; Sutton et al., 2011; Uliaszek et al. 2009; Uliaszek et al., 2010; Uliaszek et al., 2012; Vrshek-Schallhorn et al., 2011; Vrshek-Schallhorn et al., 2013; Vrshek-Schallhorn et al., 2014; Waters et al., 2014; Wolitzky-Taylor, Bobova, Zinbarg, Mineka & Craske, 2012; Wolitzky-Taylor et al., 2014; Zinbarg et al., 2010).
The Structured Clinical Interview for DSM-IV, Non-patient Edition (SCID-I/NP; First, Spitzer, Gibbon & Williams, 2002) was used to assess for DSM-IV psychiatric diagnoses. Interviews were conducted at the baseline assessment and then every 10–18 months over the subsequent 3 years. All interviewers had at least a bachelor’s degree and underwent extensive training and supervision, and interviewers presented each completed SCID at a diagnostic consensus meeting led by a doctoral-level supervisor.
Reliability for diagnoses at baseline was assessed by having trained interviewers observe live SCIDs for 69 cases. Reliability for diagnoses at follow-up assessments was assessed by having trained interviewers listen to a random selection of audio-recorded SCIDs from both sites, including at least 10% of SCIDs for each time point at each site.
Given the relatively small number of participants meeting criteria for initial onsets of many of the individual diagnoses, we conducted our primary tests at the level of diagnostic spectra. By diagnostic spectra, we mean groups of disorders classified together in the DSM-IV: UMDs included major depressive disorder, dysthymia, and depressive disorder not otherwise specified; ADs included panic disorder, generalized anxiety disorder, social anxiety disorder, obsessive-compulsive disorder, specific phobias, posttraumatic stress disorder, acute stress disorder and anxiety disorder not otherwise specified; and SUDs included alcohol abuse, alcohol dependence, non-alcohol substance abuse and non-alcohol substance dependence. In addition, we conducted separate analyses of major depressive disorder given the sufficiently large sub-sample with initial onsets, and we conducted separate analyses of initial onsets of panic disorder given its strong theoretical link with AS. Table 1 shows the new onsets of each disorder over the course of the three-year follow-up (FU) period.
When interpreting kappa (κ) values, it is important to keep in mind that κ is attenuated when the simple probabilities of the categories of a coding system deviate markedly from equiprobable (e.g., Bakeman, Quera, McArthur & Robinson, 1997). Given marked deviation from equiprobable categories in the current study due to low base rates of many disorders, we followed the recommendations of Byrt, Bishop and Carlin (1993) and Sim and Wright (2005) and report adjusted κ that adjusts for the low base rates.^3^ Adjusted κ for MDD equaled .91 at baseline, .94 at 1^st^ FU, .92 at 2^nd^ FU and .86 at 3^rd^ FU; for UMDs equaled .82 at baseline, .94 at 1^st^ FU, .88 at 2^nd^ FU, and .90 at 3^rd^ FU; for ADs equaled .76 at baseline, .85 at 1^st^ FU, .80 at 2^nd^ FU and . 76 at 3^rd^ FU; and for SUDs equaled .97 at baseline, 1.00 at 1^st^ FU, .88 at 2^nd^ FU and .83 at 3^rd^ FU. We did not have a sufficient number of cases of panic disorder in the reliability sub-samples at any time point to calculate κ. Thus overall, in the context of the low base rates, there was acceptable to very good inter-rater reliability.
At baseline, participants completed the following eight vulnerability (a) Eysenck Personality Questionnaire-Revised, Neuroticism Scale (EPQ-R-N; Eysenck & Eysenck, 1975); (b) the N scale from the International Personality Item Pool-NEO-PI-R (Goldberg, 1999); (c) The Behavioral Inhibition Scale (BIS; Carver & White, 1994); (d) the N scale from the Big Five Mini-Markers Scale (Saucier, 1994); (e) the Cognitive Style Questionnaire (CSQ; Alloy et al., 2000; Hankin, Abramson, Miller & Haeffel, 2004); (f) the Dysfunctional Attitudes Scale (DAS; Weissman & Beck, 1978); (g) the Personal Style Inventory (PSI-II; Robins Ladd, Welkowitz, Blaney et al., 1994*); and (h)* the Anxiety Sensitivity Index Expanded Form. (ASI-X; Li & Zinbarg, 2007; Reiss, Peterson, Gursky & McNally, 1986).
Participants were contacted by phone or e-mail 10 months after each SCID to schedule the subsequent SCID, and the interval between successive SCIDs was 10 to 18 months. Participants who could not be reached or were unable to complete a particular FU assessment in that time frame were contacted for the subsequent FU assessment; in all cases the FU SCIDs covered the entire period since the last completed SCID. Participants were mailed a check after completion of each assessment. All study procedures were approved by Institutional Review Boards at Northwestern University and UCLA. Of the 627 participants who completed the baseline assessment, 496 (79.1%) completed the first FU, 420 (67.0%) completed the second FU, and 422 (67.3%) completed the third FU. Of the 627 participants who completed the baseline assessment, 547 (87.2%) completed at least one of the three FU assessments and were included in the present analyses, 474 (75.6%) completed at least two, and 319 (50.9%) completed all three.
All analyses were conducted using Mplus (Muthen & Muthen, 1998 – 2012). Missing data were accommodated using full information maximum likelihood. The level of statistical significance in all inferential analyses was p ≤ .05, unless otherwise specified. There are different approaches one could take to teasing apart the unique effects of the cognitive vulnerability factors and the general N factor including conventional multiple regression. However, because we also tested predictions regarding the latent structure of the set of risk measures included in this study, we chose to use structural equation modeling. More specifically, we used the hierarchical factor model for its known strengths in separating common and unique variance sources (e.g., Chen et al., 2006; Reise, Morizot & Hays, 2007; Zinbarg et al., 2005). Thus, we began by specifying a hierarchical confirmatory factor analysis (CFA) model of N in which Dysfunctional Attitudes, Sociotropy, Autonomy, Negative Inferential Style and AS are conceptualized as N facets.^4^ We then used this hierarchical CFA model to test hypotheses regarding the (1) overlap among the risk factors and (2) unique prospective associations of the risk factors with the initial onset of UMDs, ADs and SUDs over a three-year follow-up period. In addition, we tested the invariance of the CFA model of N across males and females and the role of gender in moderating the prospective associations of the general N factor with initial onsets of UMDs, ADs and SUDs.
We randomly selected one half of the sample with which to conduct preliminary analyses including item-level exploratory factor analyses (EFAs) and initial subscale-level CFAs. To minimize capitalizing on sampling error that can arise from the use of EFA and modification indices in the model specification process, we conducted the EFAs, used modification indices and made other adjustments to models that had inadequate fit only in the first half of the sample. We then conducted confirmatory model testing in the second half of the sample (i.e., testing the models that were specified, in part, on the basis of the results in the first half of the sample). We also conducted analyses of metric and configural invariance between the two sub-samples in our final model as a further test of the extent to which we capitalized on sampling error in the model specification process.
The following fit indices were used to evaluate model fit in the CFAs: Comparative Fit Index (CFI), Root Mean Square Error of Approximation (RMSEA), and Standardized Root Mean Residual (SRMR). Hu and Bentler (1998, 1999) recommended that good fit is indicated by CFI ≥ 0.95, RMSEA ≤ 0.06, and SRMR ≤ 0.08. However, we were not rigid in our use of these cut-offs for two reasons. First, Hu and Bentler cautioned against interpreting their results as universal golden rules (see also Marsh, Hau & Wen, 2004). Second, we used three indices (which is quite common) despite the absence of data on the performance of cutoffs when using more than a pair of indices. In addition, although there are no cutoffs for the Bayesian Information Criterion (BIC), it has been shown to be useful for model comparisons (e.g., Markon & Krueger, 2006) and so we also report the BIC values for our models (with lower BIC values indicating better fit).
Proportional hazard survival analyses were conducted using a person-year database with the diagnostic variables (i.e., UMDs, ADs, comorbid UMDs and ADs, MDD, and panic disorder) as dependent variables. Individuals with a lifetime history of a particular disorder at baseline were excluded from analyses of that diagnostic outcome (for comorbid UMDs and ADs, an individual was excluded from analyses only if he or she had a history of a comorbid UMD and AD at baseline). Similarly, for individuals who developed an initial onset of a particular disorder at a given FU assessment, their subsequent person-years were excluded from the analyses of that disorder (for comorbid UMDs and ADs, an individual’s subsequent person-years were excluded from analyses only after they had comorbid UMD and AD).
We first report associations of each disorder with each of the observed measures of neuroticism and its facets. Next, we report the associations of each disorder with each of the latent variables in our vulnerability measurement model. Given that the latent variables in our vulnerability measurement model are constrained to be orthogonal, these results should be interpreted as unique effects (S. G. West, personal communication, January 25, 2013). We also conducted analyses of Pure UMDs and Pure MDD in which the outcomes being predicted were the development of new onsets of these disorders in the absence of a history of an AD. Similarly, we conducted analyses of Pure ADs in which the outcome being predicted was the development of new onsets of ADs in the absence of a history of a UMD. We did not analyze pure panic disorder given the small number of panic disorder cases and even smaller number of pure panic disorder.
Gender was a covariate in all analyses given that UMDs and most ADs are more common in women than men (e.g., Craske, 2003; Nolen-Hoeksema & Hilt, 2009) and women score higher than men on the general N factor (e.g., Costa, Terracciano & McCrae, 2001), as well as on several of our N facets (e.g., Hankin & Abramson, 2001). (Results without using gender as a covariate produced virtually identical results and are available upon request from REZ). To test whether gender moderated the associations of the general N factor with initial onsets of any of the diagnostic outcomes, we conducted multiple-group survival analyses with the general N factor as the predictor. These analyses used a likelihood-ratio test to compare a model that constrained the hazard ratio for the general N factor to be equal across the sexes versus a model that allowed that hazard ratio to differ across the sexes.
For each latent variable in the vulnerability measurement model that had a significant hazard ratio with UMDs or MDD, we conducted specificity comparisons testing whether that hazard ratio was significantly stronger than the hazard ratio for that latent variable’s prediction of ADs. Similarly, for each latent variable that had a significant hazard ratio with Pure UMDs or Pure MDD, we conducted specificity comparisons testing whether that ratio was significantly stronger than the ratio for that latent variable’s prediction of Pure ADs. Likewise, for each latent variable that had a significant hazard ratio with ADs, Pure ADs or PD, we tested whether that ratio was significantly stronger than that latent variable’s ratio with UMDs. For each latent variable that had a significant hazard ratio with Comorbid ADs and UMDs, we tested whether that ratio was significantly stronger than that latent variable’s ratio with pure cases (Pure UMDs and Pure ADs). Finally, for each latent variable that had a significant hazard ratio with SUDs, we tested whether that hazard ratio was significantly different than the hazard ratio for that latent variable’s prediction of UMDs and of ADs. We conducted all of these specificity comparisons using inferential confidence intervals (e.g., Tryon, 2001).
Preliminary analyses were directed toward specification of a base, hierarchical CFA model that could provide the foundation for testing our three hypotheses regarding the latent structure of the observed vulnerability measures. This model was specified on the basis of three considerations. The first consideration was prior theoretical and empirical research on the structure of one or more of the vulnerability measures (e.g., Lewis, Zinbarg, Craske, Mineka, Vrsheck-Schallhorn & Waters, 2010; Prenoveau, Zinbarg, Craske, Mineka, Griffith & Rose, 2009; Whisman & Friedman, 1998). The second consideration was the results of item-level EFAs conducted in the first random subsample of some of the measures. The third consideration was the results of initial subscale-level CFAs that we conducted. Details of the model specification process are presented in the Supplemental Material available online.
The preliminary analyses resulted in the identification of 15 subscales derived from the EPQ-R-N, the IPIP-NEO-PI-R-N, the Big 5 N and the BIS scales (Table S1 in the Supplemental Material available online presents the items assigned to each subscale and subscale reliabilities). Table 2 displays the correlations among those 15 subscales, as well as the correlations with the DAS-A, PSI, CSQ and AS subscales. As can be seen in Table 2, all of these correlations are positive – consistent with the presence of a general N factor (GNF) that runs through all of the subscales.
The base, hierarchical CFA model that was specified, in part, on the basis of the preliminary analyses in the first random subsample is displayed in Figure 1. All subscales were specified as indicators of a GNF. The four widely recognized N facets of depression, anxiety, self-consciousness and anger were each indicated by at least two subscales. A broad anxiety factor identified in the IPIP-NEO-PI-R by Uliaszek et al. (2009) was indicated by the IPIP-NEO-PIR-R anxiety, self-consciousness, and vulnerability subscales. BIS anxiety and self-consciousness subscales were allowed to correlate (thus accounting for the method variance due to these two subscales coming from the same measure which differed from the measure the other three subscales came from). A Sociotropy facet was indicated by the PSI-Soctiotropy and the DAS-A-Need for Approval subscales; an Autonomy facet was indicated by the PSI-Autonomy and the DAS-A-Need for Achievement subscales. A Negative Inferential Style facet was indicated by the four CSQ subscales. AS-Physical Concerns, and AS-Mental Incapacitation Concerns and AS-Social Concerns facets were each indicated by two ASI-X subscales with all six ASI-X subscales being specified as indicators of a general AS factor.
The fit of the base, hierarchical CFA model in the second random subsample was acceptable, χ^2^(339) = 766.41, p < .001, RMSEA = 0.064, CFI = 0.90, SRMR = 0.056, BIC = 16666.38. All loadings except for that of PSI-sociotropy on the anger facet were significant. We also tested invariance across the two random subsamples by conducting a multiple group analysis of our base, hierarchical CFA model. A configural invariant model provided an adequate fit to the data, χ^2^(693) = 1406.64, p < .001, RMSEA = 0.058, CFI = 0.91, SRMR = 0.054, BIC = 33208.56. In the metric invariant model, we applied across-samples equality constraints to factor loadings and intercepts for all factors and all items. The fit of the metric invariant model was adequate, χ^2^(757) = 1467.16, p < .001, RMSEA = 0.056, CFI = 0.91, SRMR = 0.061, BIC = 32858.92, and was not significantly worse than the configural invariant model, χ^2^diff(64) = 60.52, p > .10. In addition, all loadings were in the expected direction and significant. Because the metric invariant model did not significantly degrade model fit and provided adequate fit, these results indicated that the model built in the first subsample was cross-validated in the second subsample. Table 3 displays the standardized loadings from the base, hierarchical CFA model estimated in the full sample (fit in the full sample was adequate, χ^2^(339) = 973.99, p < .001, RMSEA = 0.056, CFI = 0.92, SRMR = 0.046, BIC = 32628.70).
We began our testing of the hypothesis that dysfunctional attitudes, sociotropy and autonomy are best represented by two group factors in addition to the GNF by allowing the Sociotropy and Autonomy factors to correlate. If adding this correlation provided a significant increment in model fit, it would indicate that there is additional shared variance between the Sociotropy subscales and the Autonomy subscales beyond what can be accounted for in the base, hierarchical CFA model. The fit of this model was not significantly better than the base, hierarchical CFA model, χ^2^diff(1) = 3.28, ns. Thus, the results failed to provide evidence that an additional factor beyond the GNF and the Sociotropy and Autonomy group factors is needed to model the covariances among the four DAS and PSI subscales.
We next tested whether the GNF accounts for a significant portion of the covariances among the Sociotropy and Autonomy subscales by removing the loadings of the four DAS and PSI subscales on the GNF. The fit of this model was significantly worse than the base, hierarchical CFA model, χ^2^diff(4) = 262.86, p < .001. Allowing the Sociotropy and Autonomy factors to correlate in the version of the model in which the DAS-A and PSI subscales were not specified as indicators of the GNF led to a significant improvement in model fit, χ^2^diff(1) = 36.37, p < .001. This pair of results reveals that the subscales loading on the Sociotropy factor share significant variance with the subscales loading on the Autonomy factor and that the GNF accounts for a signification proportion of this shared variance. The model with correlated Sociotropy and Autonomy factors in which the four DAS and PSI subscales were not specified as indicators of the GNF stands in a nested relationship with a useful comparison model. In this comparison model, the four DAS and PSI subscales were again not specified as indicators of the GNF but these four subscales were specified as loading on a single group factor rather than on two group factors. Comparing these two models tests whether two group factors are necessary to account for the covariances among the DAS-A and PSI subscales. Reducing the number of DAS-A and PSI group factors from two to one resulted in a significant decrement in fit, χ^2^diff(1) = 8.62, p < .001. This result indicates that the Sociotropy and Autonomy factors, although correlated, are distinguishable.
Finally, we tested whether Sociotropy and Autonomy are distinguishable from the GNF by removing the Sociotropy and Autonomy factors from the base, hierarchical CFA model. Doing so produced a significant decrement in fit, χ^2^diff(2) = 22.46, p < .001, χ^2^(341) = 788.87, p < .001, RMSEA = 0.065, CFI = 0.90, SRMR = 0.056, BIC = 16677.37.^5^ These results suggest that Sociotropy and Autonomy are distinguishable from the GNF. Thus, the pattern of results presented in this section show that the four DAS-A and PSI subscales are best modeled by two distinguishable group factors in addition to the GNF.
We attempted to test whether a factor common to all of the cognitive and personality/cognitive style vulnerability measures should be added to the model. Thus, we added one more common latent factor with the subscales of the DAS, PSI and CSQ all allowed to have loadings on this additional latent factor in addition to their other loadings described above. This model did not converge, even as we increased the number of computational iterations to 10,000. We also tried to model a factor common to the cognitive and personality/cognitive style measures after removing the sociotropy and autonomy factors, but this model also did not converge. The lack of convergence of these models suggests that the variance that the DAS, PSI and CSQ subscales all share is due to the GNF. That is, these results fail to support the existence of a distinct vulnerability factor shared by the DAS, PSI and CSQ subscales independent of the GNF.^6^
We tested whether the GNF accounts for a significant portion of the covariances among the AS subscales by removing the loadings of the six ASI-X subscales on the GNF. The fit of this model was significantly worse than the base, hierarchical CFA model, χ^2^diff(6) = 121.32, p < .001. Thus, the ASI-X subscales are indicators of the GNF. We then tested whether all six ASI-X subscales have a AS factor in common that is distinguishable from the GNF by removing the latent AS factor from the base, hierarchical CFA model. Doing so led to a significant decrement in fit, χ^2^diff(6) =157.89, p < .001. This result shows that the common AS factor is distinguishable from the GNF. Thus, the results show that the ASI-X subscales are best modeled by a common AS factor in addition to the GNF.^7^
We performed multiple group CFAs to test the invariance across men and women of the base, hierarchical CFA model of N and its cognitive facets. We began by testing for metric invariance. In the metric invariant model, we constrained each of the nonzero loadings on the various factors and each item intercept to be equal across men and women. The metric invariant model provided an adequate fit (χ^2^(757) = 1436.208, p < .001, RMSEA = 0.054, CFI = 0.91, SRMR = 0.059). We then tested whether allowing any of the factor loadings to be free to vary between men and women provided a significant increment in fit compared with the metric invariant model. None of these tests provided support for differences in loadings between men and women.^8^ Together with the adequate fit of the metric invariant model, these results suggest that the base, hierarchical CFA model of N and its cognitive facets was highly similar in men and women.
As shown in Table 4, the observed measures of N and its facets showed consistently positive associations with the development of UMDs, ADs, MDD, and Comorbid UMDs and ADs. All 48 of these hazard ratios were greater than one, with 41 (85.4%) of them being significant. In contrast, there were less consistent associations of the observed measures of N and its facets with the development of Pure UMDs, Pure ADs, panic disorder and SUDs. Only eight (16.7%) of these 48 hazard ratios were significantly greater than one, with another nine (18.8%) having point estimates less than one (though none were significantly less than one). The pattern for Pure MDD was intermediate between these first two patterns with all 12 of the hazard ratios being greater than one and with six (50%) of them being significant. The CSQ and DAS-A-Need for Achievement subscale, which have been hypothesized to be specific predictors of UMDs, had significant hazard ratios with ADs. Similarly, the ASI-X physical concerns and ASI-X social concerns subscales, which have been hypothesized to be specific predictors of ADs, had significant HRs with UMDs. Thus, none of these measures may be as specific as some have thought, but rather all of them are saturated to a substantial degree with variance due to the general N factor.
Table 4 also shows that UMDs were significantly more common in women and SUDs were significantly more common in men. None of the other hazard ratios for gender were significant. The hazard ratios for Comorbid UMDs and ADs, Pure UMDs, MDD, and Pure MDD were, however, in the direction of being (non-significantly) more common in women.
As shown in Table 5, the general N factor predicted significantly greater rates of developing each of the diagnostic outcomes except for Pure UMDs, Pure ADs and panic disorder. The depression facet predicted significantly greater rates of developing UMDs, MDD and Comorbid ADs and UMDs. The anxiety facet predicted significantly lower rates of developing Pure UMDs and Pure MDD. The inferential style facet predicted a significantly greater rate of developing Pure UMDs (but not Pure MDD). Finally, the ASI-X- mental incapacitation concerns facet predicted a significantly lower rate of developing Pure MDD and a significantly greater rate of developing panic disorder.
Only four of the comparisons testing for differences in a latent variable’s associations with UMDs versus its associations with ADs were significant. The hazard ratio for the anxiety facet was significantly smaller for Pure UMDs than for Pure ADs or even any ADs. In addition, the hazard ratio for the ASI-X- mental incapacitation concerns facet was significantly larger for panic disorder than for UMDs or major depressive disorder. None of the remaining latent variables had significantly different hazard ratios with ADs than with UMDs, including that the Inferential Style facet with Pure UMDs was not significantly larger than for any other outcome (including Pure ADs or ADs). Thus, there was very little evidence of the latent variables being significantly stronger predictors of UMDs than ADs or vice versa.
The hazard ratio for the general N factor was significantly stronger for comorbid UMDs and ADs than for ADs, Pure UMDs, Pure ADs, Pure MDD and for panic disorder. The hazard ratios for the anxiety facet were significantly smaller for both Pure UMDs and Pure MDD than for Comorbid ADs and UMDs. None of the remaining latent variables had significantly different hazard ratios with Comorbid ADs and UMDs than with the other UMD and AD outcomes.
As shown in Table 5, the general N factor and the anger facet predicted significantly greater rates of developing SUDs whereas none of the other latent variables did.
For each latent variable that had a significant hazard ratio with UMDs and/or ADs, we also conducted specificity comparisons testing whether that ratio was significantly stronger than that latent variable’s hazard ratio with SUDs. Consistent with reinforcement sensitivity theory and disconfirming the non-specificity conceptualization of N, the hazard ratios of the general N factor with UMDs, comorbid UMDs and ADs, and MDD were significantly stronger than its hazard ratio with SUDs. None of the remaining associations of a latent variable with UMDs or ADs were significantly stronger than for SUDs.
As shown in Table 5, only one of the prospective associations of the general N factor, the one with panic disorder, was significantly moderated by gender. Thus, the general N factor was associated with a significant increase in risk of initial onsets of panic disorder in females but was associated with a non-significant decrease in risk among males. The hazard ratios were also consistent with UMDs, ADs, Comorbid UMDs and ADs, Pure UMDs, Pure ADs, MDD, and Pure MDD being more strongly predicted by the general N factor in women than in men though these ratios did not significantly differ across the genders.
Our results produced five major sets of findings. First are the findings regarding N. The general N factor was a significant prospective predictor of new onsets of UMDs including MDD, ADs, Comorbid UMDs and ADs, Pure MDD and SUDs. Importantly, however, the general N factor predicted UMDs and ADs even more strongly than SUDs and predicted comorbid UMDs and ADs even more strongly. Second are the findings regarding Inferential Style and Dysfunctional Attitudes. The Inferential Style group factor was a significant predictor of Pure UMDs and did not significantly predict ADs or Pure ADs. However, the Inferential Style group factor did not predict Pure UMDs significantly more strongly than ADs or even pure ADs. In addition, our CFA results suggest that the variance shared by the DAS and CSQ subscales is attributable to the general N factor. Third are the findings regarding AS. The AS-mental incapacitation concerns facet was a significantly stronger predictor of panic disorder than of UMDs or of MDD. Our CFA results demonstrated that the hierarchical structure of N includes an intermediate-breadth AS factor in addition to the general (broad) N factor and three AS group (narrow) factors Fourth, there was little evidence of gender moderation of the ability of the general N factor to predict disorders. Finally, our CFA revealed the base, hierarchical factor model of N and its cognitive facets to be quite similar in men and women.
The results regarding N clearly disconfirm the non-specificity model advanced by Claridge and Davis (2001) of the predictive power of the general N factor (according to which, the general N factor is incapable of discriminating risk for UMDs and ADs from risk for SUDs). Rather, the results are consistent with the RST hypothesis that the general N factor is more specifically related to elevated negative affectivity, sensitivity to aversive cues and behavioral inhibition than to elevated positive affectivity, cues for reward and behavioral disinhibition (e.g., Gray & McNaughton, 2000; Zinbarg & Yoon, 2008). Thus, the general N factor was a stronger predictor of comorbid UMDs and ADs than SUDs, ADs, Pure UMDs, Pure ADs, Pure MDD and PD. This is consistent with the notion that the general N factor is especially strongly associated with comorbid UMDs and ADs and thus a core vulnerability factor common to both UMDs and ADs.
The results are partially consistent with cognitive models of vulnerability for depression. The evidence reported here for the specificity of the Inferential Style group factor as a predictor of Pure UMDs is suggestive but not conclusive given that the Inferential Style group factor did not predict Pure UMDs significantly more strongly than ADs. In addition, our results suggesting that the variance shared by the DAS and CSQ subscales is attributable to the general N factor calls into question whether the results in past research defining cognitive risk on the basis of elevations on both the DAS-A and the CSQ are attributable to the general N factor or to the Inferential Style group factor.
The evidence reported here for AS theory is also best characterized as suggestive given that the finding that the AS-mental incapacitation concerns facet was a specific predictor of panic disorder are based on only eight new onsets of panic disorder. In addition, our CFA results showed that a common AS factor can be reliably distinguished from the general N factor but our results failed to support the validity of the common AS factor or the AS-Physical Concerns or AS-Social Concerns group factors as predictors of ADs. Our results also bear on the incremental validity of negative inferential style versus dysfunctional attitudes. Our results provide support for the validity of the negative inferential style facet as a predictor of Pure UMDs. However, our results failed to support the validity of the dysfunctional attitudes facets as predictors of UMDs. This pattern of results suggests that researchers or clinicians interested in cognitive vulnerability for depression who choose to administer just one of these measures would be better off measuring negative inferential style than dysfunctional attitudes.
The current work has a number of limitations. First, selecting participants based on total scores on a screening measure for neuroticism, the EPQ-R-N, might have increased statistical power to detect unique effects of the general N factor relative to the N facets. However, simulations suggest that this is not the case (Hauner, Zinbarg & Revelle, 2014). Another limitation is that our sample has not yet entered the peak age of panic disorder onset (e.g., Kessler et al., 2006) and, as noted above, included only eight new onsets of panic disorder Additionally, we did not include measures of life stress in the analyses reported here, and several of the vulnerabilities tested here have been explicitly proposed to be diatheses that are activated by stressors (and in some cases to be activated only by congruent stressors such as sociotropy being activated by interpersonal rejection and autonomy being activated by achievement-related stressors). Finally, we did not specifically assess for hopelessness depression (Abramson et al., 1989). It is possible that if we had assessed for it, then the cognitive facets - and not the general N factor - would have uniquely predicted hopelessness depression.
In terms of future research studies to follow up on our results, one very important follow-up study would involve the design and testing of broad-based preventive interventions for high N youth. Just as trans-diagnostic treatment programs that have emerged in recent years have great potential to treat both UMDs and ADs (e.g., Barlow, Allen & Choate, 2004; Craske, 2012), our results suggest that an effective, broad-based prevention program with high N youth should hold great promise to reduce risk for both UMDs and many ADs (and especially comorbid UMDs and ADs). Our results suggest that such a preventive intervention could possibly even reduce risk for SUDs though to a lesser extent than UMDs and ADs.
Computer programs and smart-phone apps would seem to have the potential for reaching the largest number of youth. Certainly, such programs or apps could reach many of those without access to a local mental health worker. It might even be that such interventions would be acceptable to the large numbers of individuals who might benefit from mental health services and otherwise could be seen by a mental health worker but who want to solve their problems on their own (Mojtabai et al., 2011). In addition, this automated approach could be implemented in a very consistent manner without the need for training therapists. What is unclear is whether it would be more effective to attempt to directly reduce general sensitivity to threat or to enhance general emotional regulation to buffer the effects of elevated threat sensitivity. A promising example of the former strategy would be a cognitive bias modification (CBM) program to reduce attentional bias toward threat at a relatively automatic level (e.g., MacLeod, Rutherford, Campbell, Ebsworth & Holker, 2002; Schmidt, Richey, Buckner & Timpano, 2009). And an example of the latter strategy that has promise would be a CBM program to increase down-regulating cognitive reappraisal of aversive events or negative cognitions at a strategic level of processing (e.g., Denny & Ochsner, 2014; Mashal, Paller & Zinbarg, 2015). Thus, the follow-up study we most want to see conducted would randomize high N youth to one of three CBM to reduce attentional bias toward threat, CBM to strengthen down-regulating cognitive reappraisal tendencies and a control condition such as watchful waiting (e.g., Meredith, Cheng, Hickey & Dwight-Johnson, 2007).
A second important future follow-up study would be one designed to more directly elucidate the mechanisms through which N confers risk for emotional disorders. Given that the general N factor was a stronger predictor of comorbid UMDs and ADs than SUDs, we inferred that the general N factor is more specifically related to elevated negative affectivity, sensitivity to aversive cues and behavioral inhibition than to elevated positive affectivity, cues for reward and behavioral disinhibition. It would be important, however, for future research to test this notion more directly by measuring sensitivity to aversive cues and sensitivity to appetitive cues. Multiple methods should be used to assess these sensitivities including both behavioral measures and patterns of activation in threat- and reward-related neural circuitries. Such a study should yield insights that would inform the development of broad-based prevention programs. Indeed, these two research directions should ultimately converge. Full understanding of a broad-based preventive intervention requires identification of the intervention’s mechanisms just as is the case for any intervention (e.g., Kazdin, 2007) and strong causal inference about risk requires studies that manipulate the hypothesized mechanisms of risk.