The Disabilities of the Arm, Shoulder, and Hand (DASH) questionnaire is a global outcome tool for assessing the effect of upper-extremity musculoskeletal injury or condition from the patient’s perspective (Changulani, Okonkwo, Keswani, & Kalairajah, 2008; Hudak, Amadio, & Bombardier; Solway, Beaton, McConnell, & Bombardier, 2002; The Upper Extremity Collaborative Group, 1996). The DASH is intended to capture a broad picture of disability across a range of conditions affecting the shoulder, elbow, or hand (Hudak et al., 1996). The World Health Organization’s (2001) International Classification of Functioning, Disability and Health (ICF) guided the instrument’s development so that items address body function impairment, activity limitations, and participation restriction (Dixon, Johnston, McQueen, & Court-Brown, 2008; Silva Drummond, Ferreira Sampaio, Cotta Mancini, Noce Kirkwood, & Stamm, 2007). The DASH has been subject to extensive psychometric analyses; it has shown good reliability (Raven et al., 2008; Solway et al., 2002), validity (Gummesson, Atroshi, & Ekdahl, 2003; Navsarikar, Gladman, Husted, & Cook, 1999; SooHoo, McDonald, Seiler, & McGillivary, 2002), and responsiveness (Gay, Amadio, & Johnson, 2003; Greenslade, Mehta, Belward, & Warwick, 2004; MacDermid & Tottenham, 2004; Jester, Harth, & Germann, 2005; Kotsis & Chung, 2005) across a variety of orthopedic and neurological upper-extremity conditions (Beaton et al., 2001; Davis et al., 1999; Rosales, Delgado, & Díez de la Lastra-Bosch, 2002).

A global measure of disability with a wide range of item content has several advantages over a disease- or domain-specific assessment. Kinesiological theory asserts that the upper extremity works as a single unit in the completion of most, if not all, common functional tasks (Davis et al., 1999). Evidence suggests that it is difficult to design an assessment of functional items that assesses only a single joint or disease (Dowrick, Gabbe, Williamson, & Cameron, 2005). For instance, researchers have shown that instruments designed to measure wrist impairments could detect shoulder injuries and vice versa (Beaton et al., 2001). Therefore, to accurately assess function as it occurs in daily life, global measures may be necessary (Beaton et al., 2001; Gummesson et al., 2003).

Other researchers (Gay, Amadio, & Johnson, 2003; Greenslade et al., 2004; MacDermid, Richards, Donner, Bellamy, & Roth, 2000; MacDermid & Tottenham, 2004) contend, however, that compared with global scales, domain-specific scales may be more sensitive to a disorder under consideration and therefore may more accurately detect treatment-related clinical outcomes. These authors assert that although the benefits of using one instrument with all upper-extremity conditions may be appealing, disease-specific scales are more responsive. For example, much evidence suggests that the Boston questionnaire (Levine et al., 1993) is more responsive to changes in carpal tunnel syndrome than are global assessments (Amadio et al., 1996; Atroshi, Gummesson, Andersson, Dahlgren, & Johansson, 2000; Heybeli, Kutluhan, Demirci, Kerman, & Mumcu, 2002; Mondelli, Reale, Sicurelli, & Padua, 2000). Global assessments may be compounded by various factors and, hence, may fail to measure a single underlying trait. Thus, total scores obtained on global assessments may have questionable meaning. One solution lies in dividing global measures into various subscales.

The DASH is one example of a multifaceted set of items. Such multidimensional item sets become a problem when interpreting the total scores. That is, because the DASH produces a single score of summed item responses, its composite score may not distinguish among the traits being assessed. For example, what does it mean if a patient’s score decreases by 10 points after rehabilitation? Has he or she experienced fewer symptoms (body function impairment), improved functional status (activity–participation performance), or both? Decisions regarding the effectiveness of treatment rely on the meaning of the instrument’s score. Therefore, it is critical to examine the dimensionality of an instrument’s item set.

The DASH was developed according to a rigorous item generation, reduction, and psychometric testing methodology that included investigation of its dimensionality with principal-components analysis (Hudak et al., 1996; Solway et al., 2002). This analysis showed that one component explained 57% of the variance in the data, a finding that was interpreted to support the unidimensionality of the assessment and justified an aggregate scoring method. Bot and colleagues (2004), however, suggested that the DASH factor structure has not been adequately investigated.

Investigations of the dimensionality of commonly used instruments have led to dividing an assessment into domain-specific subscales. For instance, the Functional Assessment of Chronic Illness Therapy began as a generic Center on Outcomes, Research and Education (CORE) questionnaire called the Functional Assessment of Cancer Therapy–General (Webster, Cella, & Yost, 2003). This assessment, now in its fourth revision, has been divided into four different domains: Physical Well-Being, Social/Family Well-Being, Emotional Well-Being, and Functional Well-Being (Webster et al., 2003). Likewise, measures of everyday cognition, such as the Measurement of Everyday Cognition, have been divided into multiple domains (Farias, 2008).

Because the DASH is widely used in upper-extremity research and clinical practice, investigating its measurement properties is critical. The purpose of this study was to validate or challenge the factor structure of the DASH in a large sample of people with a wide range of upper-extremity orthopedic and neurological impairments.

We performed secondary analysis on existing DASH data collected by Focus on Therapeutic Outcomes, Inc. (FOTO) from various outpatient clinics throughout the United States. FOTO provides a patient assessment tool and aggregate data management service and has accumulated a large database of rehabilitation outcomes often used for research purposes (Dobrzykowski & Nance, 1997; Swinkels et al., 2007). FOTO is included on the Joint Commission’s list of acceptable performance measurement systems. The Joint Commission (www.jointcommission.org) accredits >18,000 organizations and programs that provide health care. Accreditation and acceptance by this commission is recognized nationwide as certification of quality.

The DASH is one of the many assessments on which FOTO collects data. Existing deidentified data from 991 people were obtained and analyzed according to a protocol approved by the University of Florida Institutional Review Board and in accordance with a confidentiality agreement between the University of Florida and FOTO. The DASH was administered at outpatient clinics throughout the United States before (admission) and after (discharge) rehabilitative treatment of various upper-extremity disorders.

The DASH consists of three scales: the 30-item assessment, and two 4-item optional modules (one with items relating to work and one with items relating to sports and performing arts). For the purposes of this study, we examined data obtained only from the 30-item assessment. Examples of these DASH items include “open a tight or new jar”; “make a bed”; arm, shoulder, or hand pain”; “weakness in the arm”; and “stiffness in the arm, shoulder, or hand.” Each item was rated on a 5-point ordinal rating scale: 1 = no difficulty/symptoms, 2 = mild difficulty/symptoms, 3 = moderate difficulty/symptoms, 4 = severe difficulty/symptoms, and 5 = unable/extreme symptoms. Traditionally, item responses are summed, converted to, and reported on a 100-point scale; higher scores indicate more disability. For the purposes of this study, the rating scale was recoded (i.e., reverse scored) so that low scores indicated disability and higher scores indicated ability. We agreed conceptually that results would be easier to interpret if higher scores indicated more ability.

The factor structure of the English version of the DASH was empirically examined in a study that suggested the presence of a single dominant trait (Solway et al., 2002). Several item content analyses, however, suggested that the instrument measures concepts consistent with two or three ICF domains (Dixon et al., 2008; Drummond et al., 2007), which may signal multidimensionality within the instrument. Because of the uncertainty surrounding the instrument’s factor structure and the absence of an explicit conceptual foundation supporting its item structure, we used exploratory factor analysis (EFA) rather than confirmatory factor analysis (CFA) as a first step in our investigation. That is, we chose to allow a factor structure to emerge from the data rather than approaching the analysis with an a priori hypothesis of its factor structure. We followed EFA with CFA to verify our results.

Admission and discharge data were randomly split; half of the admission and discharge data was used for the EFA, and the other half was used for the CFA. Because of the ordinal nature of the data, the EFA was conducted on the polychoric correlation matrix with weighted least square parameter estimates (Mplus software, Version 4.21; Muthén & Muthén, 1998–2007).

EFA identifies the number of factors, that is, latent traits, that best explains the pattern of covariance underlying a set of measured variables (e.g., item responses; Brown, 2006). An oblique rotation method, ProMAX (Hendrickson & White, 1964), was used because factors emerging from the data were expected to be correlated. This expectation was based on the extensive literature supporting the covariance of upper-limb joint movements (i.e., movements of the shoulder are interrelated with movements of the elbow; Shumway-Cook & Woollacott, 2000). Covariance of upper-limb joint movements is expected because of a biarticular muscular arrangement in which forces are transferred from one limb segment to another.

The number of factors was determined by examining the eigenvalues (an indication of the proportion of total variance accounted for by a factor), factor loadings, and interpretability of factors. Using the Kaiser criterion, only factors with eigenvalues >1 were examined (Portney & Watkins, 2009). Factor loadings revealed the structure of the data set by specifying the correlations between items and factors. For the purposes of this study, we considered factor loadings >0.40 as indicative of a moderately strong loading. The meaning of each factor was determined according to the item content of the items that loaded onto it.

On the basis of the results of the EFA, a CFA was conducted on the other half of the admission and discharge data. CFA is the measurement component of structural equation modeling; it tests how well the data fit a hypothesized factor structure (Brown, 2006). Goodness of fit can be indicated with a nonsignificant χ² test (p > .05), which indicates that the null hypothesis (i.e., that the model is a good fit to the data) cannot be rejected. The χ² test, however, is sensitive to large correlations between variables and large sample sizes (Brown, 2006); therefore, it often leads to inappropriate rejection of the hypothesized model (Kline, 2005). For this reason, we considered several other fit indexes: comparative fit index (CFI), Tucker-Lewis index (TLI), standardized root mean square residual (SRMR), and root mean square error of approximation (RMSEA; Hu & Bentler, 1999). The CFI and TLI assess the relative improvement in fit of the tested model compared with a baseline null model that assumes no pattern of correlation among the variables. Values of CFI and TLI >.95 indicate good fit. The SRMR is a measure of the difference between the observed and predicted correlations, with values <.10 generally considered favorable. The RMSEA is a “badness-of-fit” index, which means that a value of 0 indicates the best fit. In general, RMSEA values <.10 are considered indicative of a reasonable fit (Hu & Bentler, 1999; Kline, 2005).

In addition, we examined interfactor correlations. High correlations between factors suggest that they may be measuring the same latent trait (Kline, 2005). As a test for local independence (i.e., to ensure that two items were not so strongly correlated that they were measuring the same thing), we examined residual correlations.

Investigation of the instrument’s dimensionality was complemented by applying the Andrich Rasch rating scale (polytomous) analysis (Andrich, 1978) to item sets, using the Winsteps software program (Linacre, 2005). Mean square standardized residuals (MnSq), commonly referred to as fit statistics, were examined. MnSq represents each item’s observed variance divided by the variance expected by the measurement model. Consequently, the desired value of an item MnSq is 1.0 (Wright & Stone, 1979). The acceptable MnSq criterion for unidimensionality depends on the intended purpose of the measure and the degree of rigor desired. For sample sizes near 1,000, Linacre (2003)  and Green and Franton (2002)  recommended values between 0.6 and 1.1, associated with standardized Z values <2.0. This recommendation, however, represents a very stringent criterion. Although this criterion has been recommended for large samples, other articles have recommended that with survey (self-report) data, a value of 1.4 be used to indicate high MnSq (Wright & Linacre, 1994). Low MnSq values suggest that some other factor is interacting with the main factor and that an item therefore fails to differentiate individuals or is redundant. High MnSq values indicate that responses to that item are erratic and suggests that the item does not belong with the other items on the same continuum. For the purposes of this study, we considered only items with high MnSq values (>1.4), because they represent a threat to unidimensionality.

Interitem correlations, that is, point measure correlations produced by the Rasch analysis, were examined as further evidence of how well an item works with the other items on the assessment. Interpretation of point measure correlations is debated in the literature (Allen & Yen, 1979; Wright, 1992). For the purposes of this study, we defined an adequate correlation as >.70.

Patients spent an average of 48.6 ± 32.2 days in treatment and had an average of 13.3 ± 10.0 treatment sessions. Fifty-seven percent of the sample were women. Participants had a wide range of orthopedic and neurological conditions affecting predominantly the shoulder and neck (Table 1).

All items listed in Table 2 were analyzed, and each item’s loading on each factor is presented in the table. The first three eigenvalues were >1 (20.39, 1.46, and 1.28, respectively). Three items—(1) push open a heavy door, (2) carry a shopping bag, and (3) carry a heavy object—were cross-loaded onto Factors 1 and 2. Loadings onto the first two factors are similar for the three items; however, the items appear to be conceptually related to other items loading onto Factor 1 in that they require similar whole-body forceful movements. There is a general tendency for items loading on Factor 1 to be gross motor multijoint items. Excluding items cross-loading on the first and second factors, items grouped into Factor 1 also tended to require more range of motion. Items loading on Factor 2 tended to be fine motor items. Items loading on Factor 3 tended to be symptom items. Two items—manage transportation needs and sexual activities—failed to load onto any factor.

On the basis of the EFA results, we tested the fit of a three-factor model to the discharge data. Factor 1 included the gross motor items (Items 5–15, 18, and 19). Factor 2 included the fine motor items (Items 1–4, 16, and 17). Factor 3 included the symptoms items (Items 22–30). Items 20 and 21 were excluded from the analysis because of their failure to load on any of the three factors. The χ² value was found to be 16,023.026 (degrees of freedom = 12, p = .00). The TLI (.99) and SRMR (.05) values indicated good fit of the model to the data. The CFI (.88) and RMSEA (.14) values did not reach the criterion. The factors were moderately correlated. Interfactor correlation values were as follows: Factor 1 to Factor 2, r = .87; Factor 1 to Factor 3, r = .83; Factor 2 to Factor 3, r = .85. Residual correlations were all <.20.

Rasch rating scale analysis was applied to all items together and three item groups corresponding with the three factors. That is, using admission data, all items were analyzed together, and then separate Rasch analyses were conducted on 13 gross motor items (Items 5–15, 18, and 19), 6 fine motor items (Items 1–4, 16, and 17), and 9 symptom items (Items 22–30). Items 20 and 21 were excluded from analysis of separate constructs because of their failure to load during EFA. When all items were run together, three items had high infit: Difficulty sleeping, MnSq = 1.42; Tingling, MnSq = 1.91; and Sexual Activities, MnSq = 1.62 (Table 3). Rasch-derived infit statistics for items divided into the three constructs, point measure correlations, and item measures are presented in Table 4. Within each item subset, items are displayed in order, according to the item measure. Items at the top of each subset are “hard” or least likely to be endorsed by this sample. For gross and fine motor items, difficulties were similar to what was expected; however, within the symptoms subset, items may have been influenced by frequency of occurrence. Items at the bottom of each subset are “easy” or most likely to be endorsed by this sample. When divided into the three constructs, only one item misfit (Tingling, MnSq = 1.67).

Although the DASH is a widely used outcome measure, relatively few studies of its factor structure have been carried out to date. During the original instrument development process, principal-components analysis of the polychoric correlation matrix was applied to responses by 396 participants to 30 items to determine whether the instrument should be scored as a whole or have separate symptom and functional disability subscores (Hudak et al., 1996). The analysis showed that a single factor explained 57% of the variance in the data. Addition of another factor explained an additional 6% of the data. Interpretation of the meaning of the second factor, however, was not clear, because items loaded onto both factors even after orthogonal rotation. Consequently, we rejected the two-factor model in favor of the simpler one-factor model and recommended that the instrument be scored by summing all items.

Factor analysis was applied to the instrument when translated into Taiwanese. Liang, Wang, Yao, Horng, and Hou (2004)  conducted a principal-axis factor analysis and concluded that there is support for a one-factor structure. They found that the first factor had an eigenvalue of 13.51 and explained 45% of the total variance. Loadings on the first factor ranged from .43 to .89. None of their findings were indicative of a two-factor solution.

Similar to the results obtained by Hudak et al. (1996)  and Liang et al. (2004), we found a large first eigenvalue, which explained most of the variance in the data. Indeed, the first eigenvalue in our study explained 61% of the total variance. After oblique rotations, however, not all factors had their highest loadings on the first factor. When we considered the highest loadings for each of the items, three conceptually distinct groupings emerged: (1) gross motor activities requiring whole-body movements, (2) fine motor items, and (3) symptom items. The CFA results were also inconclusive. Although TLI and SRMR values indicated a good fit of the data to this model, the CFI and RMSEA did not reach significance.

Is the DASH unidimensional or multidimensional? A unidimensional solution is certainly plausible given our results (the first factor explained 61.32% of the total variance; CFI = .89, RMSEA = .13 with three-factor solution). Equally plausible, however, is a three-factor solution. Pragmatic reasons may support dividing the DASH into three subscales. For example, the most optimal measurement tool is one consisting of items that will measure the behaviors expected to change with an intervention. Thus, an intervention intended to alter symptoms (i.e., reduce pain) would be best measured by a tool consisting of impairment-level items. Similarly, an intervention intended to alter hand function might be best measured by a tool consisting of items that measure fine motor activities. Using multiple scales, instead of one, might provide a clearer picture of patient progress from admission to discharge.

Our rationale for supporting the division of the DASH into three subscales is as follows. Clinicians or researchers interested in changes in the acute state of an upper-extremity condition or injury require a tool to measure the signs and symptoms related to inflammation. By contrast, those interested in the severity of an upper-extremity condition or injury require a tool to measure activity and participation restriction. A global scale does not allow discernment of whether one group of patients has more active disease than another group or which limbs or joints are most affected.

Item-response theory methodologies were applied to the DASH on only one occasion. Beaton, Wright, and Katz (2005)  created the Quick–DASH using an item-reduction technique in which items that misfit the Rasch model were eliminated. In their analysis, four items showed misfit (Weakness, Tingling, Sexual Activities, and Write), where misfit was defined as MnSq values >1.3. Our results were similar. Using our criterion of MnSq > 1.4, when all items were run together, three of them misfit (Difficulty Sleeping, Tingling, and Sexual Activities), whereas when divided into three subscales, only one item, Tingling, failed to fit the Rasch model. In our separate analysis, the item Sexual Activities was excluded from analysis because of its failure to load during EFA.

This study has several limitations. First, although the sample used in this study was large, it was a sample of convenience. It was, however, representative of the reality that exists in clinics today (i.e., broadly representative of patients seen in clinics throughout the country). Second, the results of this study are somewhat inconclusive in that they do not specify whether dividing the DASH into three subscales is optimal in all cases. When a more global assessment is sufficient for the situation and time or length of the test are not issues, the DASH may be best used as a whole; when a specific condition is of concern, however, use of specific items may prove most beneficial.

We do not recommend that the standard and long-standing method of scoring the DASH be replaced. Indeed, the moderate interfactor correlations suggest that the items measure different aspects of the same trait. Instead, our study may provide a way in which a therapist could choose a subset of items for assessment instead of the entire assessment. That is, when gross motor, whole body activities are problematic, the items representing this subscale might be administered. Conversely, if pain is affecting function, the items on this subscale might be solely administered.

Future research focusing on the psychometric properties of various divisions of items on the DASH representing different constructs is warranted. These studies should include various samples with differing diagnoses. Although the DASH as a whole has achieved status as being reliable, valid, and responsive, it is uncertain how division into various subscales might affect these psychometrics. In addition, other methods to investigate dimensionality of datasets, such as parallel analysis, might be conducted to strengthen support for the factor structure of the DASH.

We thank Dennis Hart, Director of Consulting and Research of FOTO, for use of the FOTO database.