Abstract
Although obsessive-compulsive (OC) symptoms are observed along four dimensions (contamination, responsibility for harm, order/symmetry, and unacceptable thoughts), the structure of the dimensions remains unclear. The current study evaluated a bifactor model of OC symptoms among those with and without obsessive-compulsive disorder (OCD). The goals were (a) to evaluate if OC symptoms should be conceptualized as unidimensional or whether distinct dimensions should be interpreted and (b) to use structural equation modeling to examine the convergence of the OC dimensions above and beyond a general dimension with related criteria. Results revealed that a bifactor model fit the data well and that OC symptoms were influenced by a general dimension and by four dimensions. Measurement invariance of the bifactor model was also supported among those with and without OCD. However, the general OC dimension accounted for only half of the variability in OC symptoms, with the remaining variability accounted for by distinct dimensions. Despite evidence of multidimensionality, the dimensions were unreliable after covarying for the general OC dimension. However, the four dimensions did significantly converge with a latent OC spectrum factor above and beyond the general OC dimension. The implications of these findings for conceptualizing the structure of OCD are discussed.
According to the Diagnostic and Statistical Manual of Mental Disorders (5th ed., DSM-5; American Psychiatric Association [APA], 2013), obsessive-compulsive disorder (OCD) is characterized by recurrent obsessions and/or compulsions that cause marked distress and interfere with daily functioning. Obsessions are described as intrusive, repetitive thoughts, images, or impulses. Compulsions, however, are purposeful, repetitive behaviors or rituals performed in an effort to relieve obsessional distress. Although this definition of OCD has informed basic and applied research that has significantly advanced treatment efforts, determining what is OCD and what is not OCD remains somewhat unclear (Abramowitz & Houts, 2002; Abramowitz & Jacoby, 2015). This is especially relevant since OCD has been moved from among the anxiety disorders in DSM-IV (4th ed., DSM-IV; APA, 1994) to a novel group of putatively overlapping conditions—“OC-related disorders” (e.g., trichotillomania, body dysmorphic disorder)—in DSM-5. This change was made on the basis of results from comorbidity and family studies, as well as studies of neurocircuitry, neurotransmitter function, and pharmacologic treatment response suggesting that OCD is distinct from other anxiety disorders (see Bartz & Hollander, 2006, for review).
Delineating the nature of OCD and where it belongs in the DSM has been complicated by the considerable heterogeneity of the disorder (Abramowitz & Jacoby, 2015). In a review of 12 factor-analytic studies involving more than 2,000 patients, symmetry/ordering, hoarding, contamination/cleaning, and obsessions/checking were found to be the most consistently extracted OCD symptom subtypes (Mataix-Cols, Rosario-Campos, & Leckman, 2005). These OC symptom dimensions appear to be marked by differences in response to treatment (Abramowitz, Franklin, Schwartz, & Furr, 2003), patterns of comorbidity (e.g., Hasler et al., 2005; Holzer, Goodman, McDougle, & Baer, 1994), gender responding (Labad et al., 2008), neural correlates (e.g., Phillips et al., 2000), genetic transmission (Hasler et al., 2007), and underlying latent structure (Haslam, Williams, Kyrios, McKay, & Taylor, 2005; Olatunji, Williams, Haslam, Abramowitz, & Tolin, 2008). However, a disproportionate amount of research aimed at identifying specific OC symptom subtypes or dimensions has focused on the Yale–Brown Obsessive Compulsive Scale (Y-BOCS) as a flawed proxy for the construct of OCD (McKay et al., 2004). Research aimed at identifying specific OC symptom subtypes or dimensions has also focused primarily on symptom presentation (i.e., obsessional/compulsion themes). Subtype models have been proposed using factor analyses that yield dimensional systems of symptom categories but not necessarily distinct subtypes (see McKay et al., 2004, for review). Although several studies have provided support for the multidimensionality of OC symptoms (see Leckman, Rauch, & Mataix-Cols, 2007, for review), a model that may fully explain the heterogeneity of the disorder remains elusive (Summerfeldt, Richter, Antony, & Swinson, 1999; Szechtman & Woody, 2004; Taylor, McKay, & Abramowitz, 2005). Furthermore, grouping OC symptoms into dimensions based solely on overt behavioral similarities (e.g., “checking”) is limiting as it ignores the function of such behaviors (Abramowitz & Jacoby, 2015).
In the absence of definitive etiological markers for OCD, more precise delineation of OC symptom dimensions may offer a fruitful point of orientation (Leckman, Bloch, & King, 2009). For example, the dimensionality of symptoms may suggest that a given theoretical model does not apply to the development of the different variants of OCD (Taylor et al., 2006). However, advancing knowledge of the dimensionality of OC symptoms has been hampered by limitations of standard self-report measures (see Abramowitz et al., 2010, for review). Although obsessions and compulsions are heterogeneous and reflect an individual’s highly idiosyncratic concerns, several measures of OC symptoms include relatively specific items assessing the severity of obsessions and compulsions. This approach is limiting because scores on such measures may be dependent on the number and types of OC symptoms present rather than on the severity (i.e., frequency, distress, duration) of such symptoms. Another limitation of currently available measures is that they often assess obsessions separately from compulsions. However, structural analyses have consistently found thematically based OC symptom dimensions that each incorporate certain types of obsessions and compulsions (e.g., Deacon & Abramowitz, 2005; Mataix-Cols et al., 2005; McKay et al., 2004). If measures are structured to assess obsessions separately from compulsions and yet OC symptoms do not naturally distill neatly into obsessions and compulsions, this may result in an erroneous understanding of the nature and structure of the dimensions of OCD. Yet another limitation of many current measures of OC symptoms is the inclusion of hoarding symptoms. Although hoarding has traditionally been conceptualized as a symptom of OCD, empirical findings have not consistently supported a relationship between hoarding and OCD (e.g., Abramowitz, Wheaton, & Storch, 2008; Grisham, Brown, Liverant, & Campbell-Sills, 2005; Saxena, 2007; Wu & Watson, 2005). Based on such findings, hoarding is now described in DSM-5 as a disorder (hoarding disorder) that is distinct from OCD (APA, 2013).
The limitations of current measures of OC symptoms have hampered efforts to appropriately examine the dimensionality of symptoms of OCD. Such efforts would benefit from the development of a more comprehensive assessment of OC symptom presentation and impairment (Grabill et al., 2008). To fill this important gap in the literature, Abramowitz et al. (2010) developed the Dimensional Obsessive-Compulsive Scale (DOCS) that is designed to (a) assess the severity of the empirically established OC symptom dimensions in clinical and unselected populations, (b) measure symptom severity as a function of multiple empirically supported parameters, (c) assess symptom severity independent of the number or types of different obsessions and compulsions, and (d) be fairly brief and easy to administer. The DOCS assesses the four most consistently identified dimensions in previous structural analyses of OC symptoms: (a) “contamination” (contamination obsessions and decontamination [washing and cleaning] compulsions); (b) “responsibility for harm, injury, or bad luck” (obsessions about causing harm by various means and checking, reassurance seeking, and related compulsions); (c) “unacceptable obsessional thoughts” (violent, sexual, and religious obsessions with mental rituals and other forms of neutralizing); and (d) “symmetry, completeness, and exactness” (obsessions regarding something not being “just right” and compulsions involving ordering and repeating; e.g., Mataix-Cols et al., 2005). Factorial validity of the DOCS was supported by exploratory and confirmatory factor analyses (CFAs), and scores on the DOCS displayed good reliability and validity, as well as sensitivity to treatment and diagnostic sensitivity. More recent cross cultural adaptations of the DOCS have also replicated the four-factor structure (e.g., López-Solà et al., 2014; Ólafsson et al., 2013).
The development of the DOCS offers a valuable opportunity to revisit the question of whether to conceptualize OC symptoms as consisting of one general dimension or multiple dimensions. Although multidimensional solutions of OC symptoms have generally been reported in the literature (see Mataix-Cols et al., 2005), there is some evidence to suggest that a general dimension should also be considered. For example, Mathews et al. (2008) found that an underlying unidimensional component to obsessionality accounted for over 90% of the variation of traditional OC symptom factors in college students, medical students, and OCD families. Genetic research has also shown that all OC symptom dimensions share variation with a latent common factor (van Grootheest, Boosma, Hettema, & Kendler, 2008). For example, Katerberg et al. (2010) found that while the symptom structure of data from a large sample of OCD patients produced a five-factor solution, a one-factor model also emerged as a possible solution. Furthermore, heritability analyses suggested that the one-factor solution was at least as heritable as the majority of the individual factors. These findings raise the questions as to whether OCD is one disorder or a family of disorders.
A common solution to not knowing whether to consider one general OC dimension or multiple dimensions has been to do both. However, Reise, Moore, and Haviland (2010) argue that this approach can be problematic because variance in multiple OC symptom dimensions can reflect two sources—the general dimension and the specific dimension. Furthermore, multiple OC dimensions often reflect variation on both general (general OC symptoms) and more specific (e.g., contamination, responsibility for harm, unacceptable thoughts, asymmetry, or ordering) constructs. Accordingly, purportedly distinct dimensions may appear reliable due to the general variance that is also measured by the distinct dimensions. In addition to the reliability of the purportedly distinct dimensions, another important issue that is often neglected in the existing literature is how much of the variance each distinct dimension accounts for and how uniquely informative they are. In a recent study, Raines et al. (2015) found that although symmetry obsessions/arranging compulsions, harm obsessions/checking compulsions, and unacceptable thoughts/neutralizing compulsions were related to both fear and distress disorder symptoms, the contamination obsessions/washing compulsions dimension of OCD was specifically related to fear disorder symptoms. Although research of this sort contributes to the literature attesting to the multidimensional nature of OCD, it is unclear if a similar pattern of findings would be observed when controlling for variation on a general OC symptoms factor that is shared by the specific factors.
The bifactor model may have utility in accounting for an OCD symptom structure that may have multiple dimensions yet be influenced largely by a single general factor (Reise, Bonifay, & Haviland, 2013; Reise et al., 2010; Reise, Morizot, & Hays, 2007). Although the bifactor model may not be appropriate for all data and psychological phenomena, the bifactor model has been found to fit psychological data well across different problem areas, sample types, and ages (e.g., Ebesutani, McLeish, Luberto, Young, & Maack, 2014; Reise, 2012; Reise et al., 2007). One notable advantage of the bifactor model is that it allows for easier interpretation and understanding of how specific content domains relate to external variables above and beyond the general factor (cf. Chen, West, & Sousa, 2006). This is because in bifactor models, the general and group factors are orthogonal to each other and thus factors compete equally to explain item variance. The variance explained by the specific group factors therefore represents the amount of item variance accounted for by that specific group factor that is not attributed to and is independent from the general factor (Reise et al., 2010). Although the bifactor model is a highly relevant approach to better understand the structure of OC symptoms, no study to date has examined the utility of this model as a framework for conceptualizing the dimensionality of OC symptoms. An important question, however, is the extent to which a bifactor model of OC symptoms is unique to patients with OCD. Although research has revealed OC-related phenomena that is unique to OCD (Morillo, Belloch, & García-Soriano, 2007), taxometric research has shown that the heterogeneous OC symptoms are present to a greater or lesser extent in all individuals (Olatunji et al., 2008). This suggests that the bifactor model may be expected to be equivalent for those with OCD relative to those without OCD. Examination of the invariance of a bifactor model of symptoms among those with and without OCD could have important implications for etiological models. Indeed, evidence of invariance of a bifactor model may imply that various symptoms of OCD may emerge from largely normal human processes.
The primary goals of the present investigation were to (a) use bifactor modeling to better understand the latent structure of OC symptoms as measured by the DOCS (including whether DOCS item response variance is due to a general OC construct vs. distinct OC dimensions) and, in turn, (b) evaluate whether OC symptoms should be considered unidimensional or whether distinct dimensions should be additionally interpreted. Research along this line is timely given that the development of the DOCS addresses many limitations of existing scales. Indeed, the availability of this new instrument provides an opportunity to break new ground with regard to further delineating the structure of OC symptoms and how it may manifest among those with and without OCD. The incremental utility of the bifactor model was examined by comparing it to a unidimensional model as well as a previously reported correlated traits model that specifies four latent OC variables but makes no attempt to place a measurement structure on the correlated latent factors (Abramowitz et al., 2010). The present study also examined invariance of the bifactor model between those with and without OCD. Structural equation modeling (SEM) was also used to examine differences in the association between a latent OC spectrum variable and a general latent OC variable relative to latent variables corresponding to the dimensions of the DOCS. An “OC spectrum” latent factor was formed that consisted of cognitive and behavioral symptom subtypes such as washing, checking/doubting, obsessing, neutralizing, ordering, and hoarding. SEM was then employed to examine whether the dimensions of the DOCS provide incremental utility above the general OC factor in corresponding with the OC spectrum latent factor. It was predicted that a general latent OC factor and latent variables corresponding to distinct OC dimensions would converge with the OC spectrum latent factor. Given that no study to date has examined the utility of a bifactor model of OC symptoms, findings from this research will have implications for further consideration of the nature and structure of OCD.
Method
Participants
The sample included a total of 1,299 participants: 246 with OCD, 158 with a psychological disorder other than OCD, and 888 unselected adults (with 7 participants with missing diagnostic information). Participants in the full sample ranged in age from 12 to 78 years to capture a broader developmental spectrum. All individuals in the OCD group were seeking psychological services at six treatment settings and all met DSM-IV text revision (APA, 2004) criteria for OCD as assessed by a structured or semistructured interview such as the Anxiety Disorders Interview Schedule (Di Nardo, Brown, & Barlow, 2004) or Y-BOCS (Goodman, Price, Rasmussen, Mazure, Delgado, et al., 1989, Goodman, Price, Rasmussen, Mazure, Fleischmann, et al., 1989). The OCD group included 136 (55.3%) females and 106 (43.1%) males, 178 (72.4%) participants were Caucasian, and their mean age was 29.00 years (SD = 11.93).
Patients in the other disorders group also presented for psychological services at various treatment settings. In this group 47 (29.7%) had a diagnosis of panic disorder, 39 (24.7%) had a diagnosis of social phobia, 30 (19.0%) had a diagnosis of generalized anxiety disorder, 11 (7.0%) had a diagnosis of specific phobia, 9 (5.7%) had a diagnosis of posttraumatic stress disorder, 2 (1.3%) had a diagnosis of body dysmporhic disorder, and 20 (12.7%) were diagnosed with psychological disorders that fell in an “other” category. 1 This group included 80 (50.6%) females and 78 (49.4%) males; 129 (81.6%) participants were Caucasian, and their mean age was 31.58 years (SD = 13.69).
The unselected group (n = 888, 68.4%) was recruited from undergraduate student populations at large universities in the southern United States. These participants were awarded course credit for participation in the present study. This unselected group included 598 (67.3%) females and 287 (32.3%) males; 659 (74.2%) participants were Caucasian, and their mean age was 19.16 years (SD = 1.91).
Procedure
The procedures for diagnosing OCD and other disorders were as follows: All individuals presenting for evaluation and treatment at the various study sites completed a packet of self-report questionnaires that included measures of OCD and related phenomena. Next, each participant was given a diagnostic interview by a trained assessor to establish the presence of OCD or another psychological disorder. Depending on the site, diagnoses were made on the basis of the Anxiety Disorders Interview Schedule (Di Nardo et al., 2004), the Mini International Neuropsychiatric Interview (Sheehan et al., 1998), Structured Clinical Interview for DSM-IV (First, Spitzer, Gibbon, & Williams, 2002), the Y-BOCS (for OCD patients only), or a standard clinical interview based on the DSM-IV. After completing the initial interview and formulating a diagnosis, the first interviewer presented the assessment data to a more expert clinician (i.e., the director or senior clinician at each site), who subsequently met and reviewed the assessment data with the patient. Although formal interrater reliability checks were not conducted at all sites, only patients for whom both interviewers agreed on diagnostic status were included in the study (i.e., 100% interrater agreement). Unselected participants completed measures of OC and related phenomena as part of a battery of self-report questionnaires.
Measures
Dimensional Obsessive-Compulsive Scale (Abramowitz et al., 2010)
This is a 20-item self-report measure that is divided into four subscales that assess the four most common OCD symptom dimensions: (a) contamination, (b) responsibility for harm, (c) unacceptable thoughts, and (d) symmetry or ordering. Because of the variability and heterogeneity of OCD symptoms within each dimension, each subscale begins with a description of the symptom dimension as well as examples of associated obsessions and rituals to clarify each dimension’s fundamental obsessional fears, compulsive rituals, and avoidant behavior. Each subscale contains five items (rated 0-4) that assess the severity of the following parameters over the past month: (a) time occupied by obsessions and rituals, (b) avoidant behavior, (c) associated distress, (d) functional interference, and (e) difficulty ignoring the obsessions and abstaining from the compulsions. The DOCS subscales have excellent reliability in clinical samples (α = .94-.96) and converge well with other measures of OC symptoms (Abramowitz et al. 2010).
Obsessive-Compulsive Inventory–Revised (OCI-R; Foa et al., 2002)
This is an 18-item measure of OCD symptoms. Individuals are asked to indicate the degree to which they are bothered or distressed by OCD symptoms in the past month on a 5-point scale, ranging from 0 = not at all to 4 = extremely. The OCI-R yields six subscales associated with different types of OCD symptoms (i.e., Washing Concerns, Checking/Doubting, Obsessing, Mental Neutralizing, Ordering, and Hoarding). These subscale scores were used as the observed indicators of the OC spectrum factor used in the SEM analyses, described below.
Data Analytic Approach
Confirmatory Factor Analysis
Prior research examining the categorical versus dimensional nature of OC symptoms using taxometric methods have supported dimensional models that suggest that such symptoms are present to a greater or lesser extent in all individuals (Olatunji et al., 2008). Based on such findings, we conducted CFA using Mplus Version 7.11 (Muthén & Muthén, 2010) to examine the underlying structure of the dimensions of OC symptoms, as assessed by the DOCS. As with the initial validation study (Abramowitz et al., 2010), we examined our data as categorical outcomes and used polychoric correlation matrices and the weighted least squares means and variance adjusted estimator when conducting analyses (Flora & Curran, 2004; Muthén, du Toit, & Spisic, 1997). The primary structures examined via these CFA analyses were (a) the originally posited four-factor correlated-traits model corresponding to Contamination, Responsibility, Unacceptable Thoughts, and Symmetry related OCD (Figure 1, Model A; Abramowitz et al., 2010), (b) the four-factor bifactor model of OC symptom dimensions (Figure 1, Model C), and (c) a parsimonious one-factor model (unidimensional model whereby all items loaded on a single OCD factor; Figure 1, Model D). Since correlated-traits models (in general) yield the same fit indices as second-order models (Figure 1, Model B; Brown, 2006), we did not report the fit of both models in Table 1. We examined and reported fit of only the four-factor correlated-trait model (and not the second-order model). The second-order model (i.e., Figure 1, Model B) is shown just for illustrative purposes. Significance of factor loadings was tested by dividing each factor loading estimate over its standard error estimate and treating the resulting value as a z-score.
Competing OCD models.
Fit Statistics for the Confirmatory Factor Analytic Models.
Note. df = degrees of freedom; CI = confidence interval; RMSEA = root mean square error of approximation; CFI = comparative fit index; TLI = Tucker–Lewis Index; AIC = Akaike information criterion; BIC = Bayesian information criterion; OCD = obsessive-compulsive disorder.
To evaluate and compare model fit, we used the root mean square error of approximation (RMSEA; Steiger, 1990), with values lower than .08 and lower than .05 used as benchmarks for adequate and good fit, respectively (Browne & Cudeck, 1993). We also used the comparative fit index (CFI; Bentler, 1990) and Tucker–Lewis index (TLI; Tucker & Lewis, 1973) to evaluate model fit, and used values greater than .90 (Bentler, 1990) and .95 (Hu & Bentler, 1999) as benchmarks for adequate and good model fit, respectively.
Given the complexity of comparing bifactor models, we also examined the Akaike information criterion (AIC; Akaike, 1973) index for each model. AIC is similar to chi-square, but it adjusts for model complexity (via a “penalty” function) and rewards model parsimony (Sakamoto, Ishiguro, & Kitagawa, 1986). Models can be compared with respect to AIC values, and the model with lower AIC is associated with better model fit (Kline, 1998). We also evaluated the Bayesian information criterion (BIC; Schwarz, 1978), which is another recommended model fit statistic that rewards parsimony and penalizes for model complexity (Kass & Raftery, 1995). Both were evaluated simultaneously when comparing between models. We also compared tests using the chi-square difference test (which we computed in Mplus via the “Difftest” command due to basing analyses on categorical data and using the weighted least squares means and variance adjusted estimator).
Measurement Invariance Analysis
We conducted measurement invariance of the DOCS bifactor model across individuals with an OCD diagnosis (n = 246) and individuals without an OCD diagnosis (n = 1,046). Here, data from the unselected participants and those with a psychological disorder other than OCD were combined to form a group of the individuals without an OCD diagnosis. This was done for two reasons. First, the relatively smaller sample size for those with a psychological disorder other than OCD would preclude a bifactor model that could be meaningfully interpreted. Second, combining the two groups allows for a more representative range of scores.
We tested measurement invariance via the multigroup CFA approach recommended by Brown (2006). This involves evaluating for various levels of measurement invariance beginning with single-sample solutions specific to each subsample. If both groups evidence good model fit, then configural invariance (“equal form”) is tested. In other words, are the factors associated with the same item clusters across groups? Support for configural invariance (whereby only the general item-to-factor structure are held equal across groups and all factor loadings and thresholds are allowed to vary across groups) is determined based on whether this “equal-form” model’s fit indices meet the previously mentioned cutoffs. If configural invariance is supported, then metric invariance and scalar invariance are tested. In other words, do the factors have the same meaning across groups and are comparisons of group means meaningful? To test this, factor loadings and item thresholds are held equivalent across groups and invariance of these parameters is assessed based on whether these constraints significantly degrade model fit. If model fit is not significantly degraded, then both groups are presumed to have generally equivalent factor loadings and item thresholds. Although the chi-square difference test has been used in the past to examine degradation in model fit, Cheung and Rensvold (2002) have argued that the chi-square difference test often inappropriately leads to the rejection of invariant models. They thus recommend using the ΔCFI < .01 criterion—the lack of substantial drop in CFI—as the criterion for the metric and scalar invariance tests. Notably, with categorical data, factor loadings and item thresholds need to be freed and constrained in tandem (because the item probability curve is influenced by both parameters, thereby rendering interpretation difficult if these parameters were to be freed and constrained in successive order; Muthén & Muthén, 2010). We therefore first examined single-sample solutions in each of the specific subsamples, then configural invariance (i.e., “equal form”) based on the full sample, and then scalar invariance based on the full sample (which included simultaneously holding factor loadings equivalent across groups).
Reliability
Bifactor models offer the ability to use alternative statistics to better estimate reliability. Such statistics include the omega indices, which are (confirmatory) “model-based” indices that take into account reliability due to the general factor. Omega indices are also based on congeneric rather than tau equivalent models (see Raykov, 1997a, 1997b, for the differences between these two model types). Congeneric models, for example, do not force items to be on the same scale or to be measured with the same degree of precision or amounts of error but rather assume that all items measure the same construct but with possibly different scales, different degrees of precision, and different amounts of error (Raykov, 1997a, 1997b). Indices (e.g., the omega indices) that are based on the congeneric model are therefore more likely to reflect the differences that exist in reality (with respect to scales, precision, and error differences). Accordingly, omega indices are considered to be better able to yield more accurate estimates of reliability (Graham, 2006).
Omega
Based on our confirmatory bifactor results, we computed OmegaGeneral (which represents the reliability of the total score based on all sources of reliable variance across all the items) and OmegaSubscale (which represents the reliability of a particular subscale based on all sources of reliable variance across the items from that subscale).
Omega hierarchical
We were also interested in the omega hierarchical statistic (OmegaH; Zinbarg, Barlow, & Brown, 1997; Zinbarg, Revelle, Yovel, & Li, 2005). This is a recommended statistic for evaluating the dimensionality of bifactor models (Revelle & Zinbarg, 2009), and it can be applied to both the total score and subscale dimensions. Omega hierarchical for the total score (OmegaHTotal) estimates the amount of total composite score variance explained by only the general dimension. This index reflects the degree to which the total score reflects a single latent construct. Omega hierarchical for each subscale (OmegaHSubscale) provides an estimate of subscale reliability, controlling for reliability due to the general factor. Naturally, Omega hierarchical will thus be lower than Omega. Although strict cut points do not exist for “good” reliability, we considered .70 as a general cut point of adequate reliability.
As a point of comparison, we also reported the omega statistics, which are akin to coefficient alpha and represent all sources of common variance. The omega statistics are thus expected to be much higher than the omega hierarchical statistics.
Explained Common Variance and Percentage of Uncontaminated Correlations
The explained common variance (ECV) and percentage of uncontaminated correlations (PUC) were computed (based on the confirmatory bifactor model), as these indices have been recommended when it is suspected that a particular set of data has a multidimensional structure that takes a bifactor form. These indices inform the consequences of forcing multidimensional data into a unidimensional structure.
The ECV statistic represents the percent of common variance that can be explained by the general factor (see Reise, Scheines, Widaman, & Haviland, 2013). ECV can be interpreted as an index of “unidimensionality” when dealing with potentially multidimensional data such as the DOCS. High ECV indicates that the general factor accounts for the majority of all common variance, whereas low ECV indicates that more item variance can be accounted for by specific factors beyond the general factor.
The PUC statistic represents the degree to which the factor strength will be biased due to forcing a bifactor structure into a unidimensional model. As PUC increases, the average level of structural coefficient bias lessens. PUC is increased when items do not cross-load and when more group factors are pure indicators of the general factor. The cutoff of .80 has been suggested as a general PUC value at which loading estimates in the unidimensional model will be adequately close to true loadings on the general factor in a bifactor model (when data are actually multidimensional but fit in a unidimensional model; Reise et al., 2013). These researchers (Reise et al., 2013) also stated that “when PUC is lower than .80, researchers may consider ECV values greater than .60 and omegaH values greater than .70 as tentative benchmarks” for when bias on parameters is expected to be low when fitting multidimensional data into a unidimensional model (p. 22).
Structural Equation Model: Unique Association With Related Criteria
Last, we used SEM to examine whether the DOCS (general and specific) OC related domains converged with an OC spectrum latent factor. The DOCS items were used to identify the general and specific OC domains (consistent with the bifactor model described and tested above); the OC spectrum latent factor was identified by the subscales of the OCI-R (i.e., Checking, Orderliness, Neutralizing, Washing, Obsessions, and Hoarding subscales). Given the orthogonal nature of the DOCS factors in the bifactor model, the OC spectrum latent factor was regressed on all DOCS (general and specific) domains simultaneously. Given the multiple simultaneous paths estimated in the SEM analyses, we used Bonferroni correction (.05/5 = .01) to adjust our p value to .01.
Results
Confirmatory Factor Analysis
One-Factor (Unidimensional) Model
All factor loadings of this unidimensional model were significant (ranging from .62 to .88). However, fit of this model was poor (e.g., RMSEA = .200, CFI = .836, TLI = .816). Fit indices appear in Table 1.
Correlated-Traits Model
The four-factor correlated-traits model (with factors corresponding to Contamination, Responsibility, Unacceptable Thoughts, and Symmetry) evidenced good model fit (see Table 1) All factor loadings were significant. The six correlations among the four factors in this model were also all significant and were as follows: Contamination with Responsibility (r = .59), Contamination with Unacceptable Thoughts (r = .39), Contamination with Symmetry (r = .53), Responsibility with Unacceptable Thoughts (r = .60), Responsibility with Symmetry (r = .60), and Unacceptable Thoughts with Symmetry (r = .48).
Bifactor Model
The four-factor bifactor model was associated with excellent model fit (i.e., RMSEA = .049, CFI = .991, TLI = .989). All items also loaded significantly on the general OC factor (ranging from .45 to .85) as well as on each of the four content domains, ranging from .39 to .73 (see Table 2 for all factor loadings).
Confirmatory Factor Analysis Factor Loadings and Standard Errors of the Dimensional Obsessive-Compulsive Scale (DOCS) Bifactor Model.
Note. All factor loadings appearing above were significant. Standard errors appear in parentheses.
Model Comparisons
Examination of all fit indices revealed strongest support for the four-factor bifactor model relative to the four-factor correlated traits model. In particular, the bifactor model was associated with smaller AIC and BIC values relative to the correlated traits model, and the chi-square value of the bifactor model was significantly lower than that of the correlated traits model (see Table 1). The parsimonious unidimensional model was also associated with poor fit indices (e.g., RMSEA = .200, CFI = .836) and was also associated with significantly worse fit (based on the chi-square difference test; see Table 1). Results thus indicate that the bifactor structure best represents the underlying structure of the DOCS items.
Measurement Invariance Analysis
Bifactor Model
Measurement invariance of the bifactor model was supported across those without an OCD diagnosis 2 and those with an OCD diagnosis. First, single-sample solutions revealed good fit for the bifactor model in the OCD subsample (RMSEA = .059, CFI = .994, TLI = .992) and the non-OCD group (RMSEA = .049, CFI = .985, TLI = .981). All fit indices may be seen in Table 1. Factor loadings on both the general factor and specific group domains were also significant in both subgroups.
Configural Invariance
Support for the configural invariance model across OCD diagnostic status was also found, as supported by strong fit indices above benchmark for good fit (i.e., RMSEA = .046, CFI = .991, TLI = .989).
Scalar Invariance
Support for scalar invariance (which held factor loadings and thresholds equivalent across groups) was also found, as evidenced by strong fit indices (RMSEA = .040, CFI = .991, TLI = .991), and ΔCFI not exceeding .01. It is worth noting, however, that the chi-square difference test was significant. In the absence of the ΔCFI test, the significant chi-square difference test would suggest the presence of some noninvariant item properties.
Reliability
Table 3 presents the reliability statistics (i.e., confirmatory bifactor model-based omega and omega hierarchical indices) for the total score and four subscales. OmegaHTotal shows that the general OC dimension accounted for a large amount of composite score variance (.81). Coefficient omega hierarchical for the subscales (which reflect subscale reliability estimates after controlling for the general factor) were as follows: Contamination = .53, Harm = .25, Thoughts = .54, and Symmetry = .45. The general OC dimension was the only dimension associated with high reliability (>.70).
Alpha, Omega, Omega Hierarchical, and Explained Common Variance Statistics Based on the Confirmatory Factor Analysis Bifactor Model.
For comparison purposes, the omega indices (which include all sources of common variance) were as follows: General = .98, Contamination =.95, Harm = .95, Thoughts = .96, and Symmetry = .96.
Explained Common Variance and Percentage of Uncontaminated Correlations
The ECV statistic associated with the confirmatory bifactor model was .54. This suggests that the general OC factor does not account for the majority of all common variance; in fact, about an equal amount of common variance appears to be explained by factors beyond the general factor.
The PUC statistic associated with the confirmatory bifactor model was .79. This suggests that although the OC items may be multidimensional, fitting them into a unidimensional structure (e.g., as in a bifactor model with a general factor) will not introduce a high degree of bias on structural coefficients.
Structural Equation Model: Unique Association With Related Criteria
The measurement model underlying the SEM model (whereby an OC spectrum latent factor was regressed on the DOCS dimensions) demonstrated acceptable fit (e.g., RMSEA = .060, CFI = .929, TLI = .915, SRMR [standardized root mean square residual] = .058). 3 SEM results of this model revealed that the general OC dimension and all specific OC domains significantly converged with the OC spectrum latent factor (see Table 4). As shown in Figure 2, the general OC factor most strongly converged with the latent OC spectrum factor (ζ = .69; SE = .058). The next largest path showing significant convergence with the latent OC spectrum factor was the DOCS Symmetry dimension (ζ = .32). These results are consistent with the ECV findings indicating that additional common variance can be explained beyond the general OC factor. As a point of comparison, an SEM model whereby the OC spectrum latent was regressed on a higher order OC factor (identified by the four factors of the four-factor correlated traits model; see Figure 1) was also examined. In this (nonbifactor) unidimensional model, the higher order OC factor did not significantly converge with the OC spectrum factor (ζ = .12; SE = .227).
DOCS Bifactor Model Predicting an Obsessive-Compulsive Spectrum Latent Factor as Assessed by the Obsessive-Compulsive Inventory–Revised: Structural Equation Modeling Measurement Model and Structural Parameters.
Note. DOCS = Dimensional Obsessive-Compulsive Scale; OCD = obsessive-compulsive disorder; CFI = comparative fit index; TLI = Tucker–Lewis index; RMSEA = root mean square error of approximation; SRMR = standardized root mean square residual.
Convergence of the four-factor bifactor model in relation to OC spectrum symptoms.
Discussion
Although there is increasing consensus that OC symptoms are heterogeneous clinical phenomena that should be assessed, diagnosed, and treated from a multidimensional perspective (Mataix-Cols et al., 2005), empirical demonstrations of this dimensionality have been hampered by the limitations of most existing self-report measures. Using a recently developed measure of OC symptom dimensions that addresses important limitations of existing measures, the present study found strong support for a bifactor model in which a general OC symptom factor coexists alongside four specific theme-based symptom dimensions. In fact, the bifactor model showed a better fit to the data than a competing four-factor correlated traits model previously reported in the literature (Abramowitz et al., 2010), despite being “penalized” statistically for its complexity. Thus, the bifactor model observed in the present investigation suggests that OC symptoms exhibit some communalities captured by a general OC factor, while also exhibiting distinct components that are separate and unique from the general OC factor. The ECV statistic also indicates that OC symptoms are associated with nonnegligible amounts of multidimensionality that can be attributed to the four distinct dimensions. It is possible that the four dimensions represent narrow factors that arise due to method effects. However, this seems unlikely given the findings of the validity analyses. Rather, these findings suggest that each of the four dimensions corresponds to a distinct set of mechanisms that may be further defined by an aggregate of causal factors that incrementally influence the risk for a particular cluster of OC symptoms (Taylor, 2004).
The present study also found that the general OC factor and the four specific OC symptom dimensions were associated with good coefficient alphas. Omega indices were also used as estimates of reliability since they generally provide better estimates than coefficient alpha and because they do not conflate reliability of the specific OC symptom dimensions with reliability attributed to the general OC factor. The omega hierarchical for the DOCS total score—which estimates the amount of total score variance associated with variation on the general OC factor common to all the items—was adequate. Omega hierarchical for each DOCS subscale, which provided an index of the degree to which the four specific OC symptom dimensions scores provide reliable variance after accounting for the general OC factor, revealed that the reliability of all four specific OC symptom dimensions dropped significantly once accounting for the general factor. One interpretation of the significant drop in the reliability of the specific factors when covarying for the general factor is that the distinct factors may share phenomenological similarities (i.e., obsessive thinking and/or compulsive behaviors) that are common to OCD in general.
The OC symptom dimension of obsessions about responsibility for harm is among the most commonly identified presentation of OCD (e.g., Mataix-Cols et al., 2005; McKay et al., 2004). However, the responsibility for harm dimension yielded the lowest Omega hierarchical relative to the other dimensions. Given this finding, researchers should consider that this DOCS subscale may offer minimal information above and beyond the general OC factor. Table 2 also showed that the responsibility for harm dimension had relatively lower factor loadings in the bifactor model relative to the other dimensions. While the dimension-specific factor loadings for contamination, unacceptable thoughts, and asymmetry or ordering dimensions were generally higher than loadings on the general OC factor, the opposite was true for the responsibility for harm dimension. This finding suggests that the responsibility for harm dimension may be less differentiated from a general OC factor than the other dimensions. The finding that the responsibility for harm dimension may have limited reliability when accounting for the general OC factor may also be due to distinct item characteristics. However, it has been proposed that elevated responsibility for causing (or failing to prevent) harm is relevant to all presentations of OCD (Salkovskis, 1996; Salkovskis & Forrester, 2002; Wilson & Chambless, 1999). Indeed, research has shown that exaggerated beliefs about responsibility play a role in the development of OCD (Arntz, Voncken, Goosen, 2007; Shafran, 1997) and are specific to OCD (Cougle, Lee, & Salkovskis, 2007). Accordingly, the finding that the reliability of the responsibility for harm OC dimension is largely accounted for by the general OC factor may suggest that responsibility for causing harm is relevant to all variants of OCD.
The present study also examined whether the fit of the bifactor model is equivalent for those with OCD relative to those without OCD. This question may be informed by prior research showing that the heterogeneous symptoms of OCD are best conceptualized as dimensional in that they are present to a greater or lesser extent in all individuals (Olatunji et al., 2008). Consistent with this view, strong evidence for measurement invariance of the bifactor model was seen across those with OCD and those without OCD in the present study. An important implication of this finding is that the structure of heterogeneous symptoms that is being measured among those with OCD is the same structure that is being measured among those without OCD. The invariant bifactor model appears to fit well with recent work documenting that OC symptoms (a) are prevalent in nonclinical populations, (b) are dimensional as opposed to categorical, (c) have similar etiologies in clinical and analogue samples, (d) are phenomenologically similar across clinical and nonclinical individuals, and (e) are associated with the same developmental and maintenance factors in clinical and analogue samples (Abramowitz et al., 2014). The invariant bifactor model would indicate that research on the structure of OC symptoms among those without an OCD diagnosis is highly relevant to understanding OC-related phenomena in individuals diagnosed with OCD.
The present study also examined convergence of the bifactor model factors with related OCD criteria. Specifically, SEM was used to examine differences in the association between a latent OC spectrum variable and a general latent OC symptom variable relative to the four latent variables corresponding to the four DOCS subscale domains. A latent OC spectrum variable was formed, and then SEM was used to test whether the DOCS subscales predict any additional variance over and above the general OC symptom factor. The findings showed that the general OC dimension and all specific OC domains significantly corresponded with the latent OC spectrum factor. This was in contrast to a higher order OC factor that did not significantly corresponded with the OC spectrum factor. Examination of the pattern of associations of the bifactor model did show that the general OC factor most strongly corresponded with the latent OC spectrum factor. However, this finding may be partially accounted for by the broad symptoms of washing concerns, checking/doubting, obsessing, mental neutralizing, ordering, and hoarding that served as indicators of the latent OC spectrum factor. Despite the robust effect of the general OC factor in corresponding with the latent OC spectrum factor, the four content dimensions did yield significant associations with the latent OC spectrum factor when covarying for the general OC dimension. This is strong evidence that additional common variance can be explained by the four content dimensions that is above and beyond the general OC factor.
Regarding scoring of the current version of the DOCS given results of the current study, researchers are advised to compute both the subscale scores and the total score for the purposes of interpretation. The validity tests of the bifactor model suggest that the DOCS subscales should not be avoided because they may provide additional information beyond the total score. However, the low reliability of the scales suggests that some revision may be warranted. The psychological assessment of any construct is a fluid process that should be informed by emerging research. To increase the reliability of the DOCS scales, more basic behavioral and psychometric research will be needed to more precisely delineate the processes that are unique to the specific OC symptom domains. Research along these lines may then inform the inclusion of items in a revised version of the DOCS that yields more distinctive and reliable OC symptoms domains.
This is the first investigation to our knowledge to examine a bifactor model of OC symptoms in a diverse sample of clinical and unselected participants. Although this represents an important contribution to the existing literature, additional studies are needed to verify whether the bifactor model best characterizes OC symptoms using additional measures that are derived from alternative views of the structure of OCD. Although the development of the DOCS has addressed many limitations of existing measures of OC symptoms (Abramowitz et al., 2010), other measures consist of different symptom compositions that may yield a different hierarchical structure. For example, a recent study using alternative measures of OC symptoms found that a three-class solution consisting of “autogenous obsessions,” “reactive OC,” and “reactive obsessions” was the best fit to data from patients with OCD (Atli, Boysan, Cetinkaya, Bulut, & Bez, 2014). This approach to categorizing OC symptoms is based largely on the observation that there are some obsessions that are highly aversive and perceived as threatening in their own right and that there are other obsessions where the perceived threat tends to be not the obsession itself but rather its possible negative consequence (Lee & Kwon, 2003). Although there has been some research supporting the classification of obsessions into subtypes that are marked by autogenous versus reactive obsessions (Lee, Kwon, Kwon, & Telch, 2005), future research is needed to examine if this distinction also fits well with the bifactor approach.
Additional limitations of the investigation are also worth noting. Although the present study employed a large sample of participants with a wide age range to increase generalizability, it remains unclear if the fit of the bifactor model would significantly differ between children and adults. Similarly, an important limitation is that the unselected sample was exclusively students. The generalizability of these findings would be enhanced with the inclusion of unselected participants from the community. Longitudinal studies suggest that OC symptoms tend to be more stable in adults compared to children (Mataix-Cols et al., 2002; Rettew, Swedo, Leonard, Lenane, &Rapoport, 1992). Replication of this research with adequate sample sizes that would allow for a meaningful comparison between children and adults may then be warranted. It would also be meaningful for future research to conduct item response theory–based differential item functioning to examine item-level measurement invariance to reveal whether bias exists between certain groups, such as OCD clinical groups, non-OCD clinical groups, and nonclinical groups. Another limitation of the present investigation is that the examination of the validity of the bifactor model was limited to other OC symptoms. It is important to note that the pattern of associations observed for the DOCS general factor and its subscales may not necessarily be the case for other criterion variables. This limitation is also relevant to the evaluation of the convergent validity of the bifactor model in relation to a latent OC spectrum factor. Indeed, the bifactor model and the latent OC spectrum factor consisted of self-report of a similar cluster of symptoms (i.e., both measures have scales relevant to contamination, checking/doubting, obsessions, and ordering). Furthermore, both were administered at the same time which precluded our ability to examine the predictive validity of the bifactor model. Future research that assesses a wider range of criterion variables will be useful in determining the incremental validity of the general OC dimension and the four content dimensions. Indeed, future research where the individual components of the bifactor model are validated using objective assessments of overt behavior would offer greater confidence in the present findings.
The present findings suggest that OC symptoms may be considered unidimensional, but distinct dimensions should also be interpreted. More specifically, these findings suggest the DOCS dimensions should be interpreted, with the possible exception of the responsibility for harm domain. Despite some limitations, the bifactor model may offer a new way of modeling the heterogeneity of OC symptoms. In applying this approach to the DOCS, the present investigation revealed that the heterogeneity of OC symptoms likely consists of a general OC factor that coexists alongside specific OC symptom dimensions. This finding should prove useful in future efforts to further refine the structure of OCD.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.