Positivum: Beliefs and perceptions measure

The PBPS is part of a more comprehensive assessment of BPS factors used within the OR context, particularly with those who have been injured and are receiving support through the workers compensation or personal injury compensation regulators. Based on the HBM, the PBPS measure was developed to explore an individual’s beliefs and perceptions and their potential interference with rehabilitation goals and, in particular, return to or commencement of work.

The initial iteration of the overarching Positivum assessment emerged based on:

• A literature review in 2015 conducted by Monash University that gathered evidence on psychosocial risk factors associated with poor work and health outcomes following injury (unpublished data)
• The findings and consensus decisions of an expert panel that included a state-based workers compensation representative, an occupational physician, Monash University academics with expertise in work-related disability and injury outcomes, and vocational rehabilitation specialists (6 persons in total, all professional contacts of the research team leading the development of this new measure)

A systematic review of standardised measures that were already available and validated was then conducted to determine the suitability of these measurement tools for assessing the key BPS areas of interest, and their suitability for this context. A number of existing questionnaires were short-listed, and the expert panel re-engaged to make decisions regarding their relevance. No existing validated scales were identified that comprehensively assessed an individual’s beliefs and perceptions that were relevant to the prevention of work disability. As such, individual items deemed to be of relevance were pulled from existing scales to be considered for inclusion in a new ‘fit for purpose’ measure. Existing scales from which items were drawn and adapted included The Keele StarT Back Tool [10], Readiness for Return to Work scale [11], The Survey of Pain Attitudes [12], and the Work-related Recovery Expectations Questionnaire [13]. These items were selected and modified where necessary by a working group emerging from the expert panel.

An initial pool of 22 items was assembled by the working group that included items belonging to each of three pre-determined, evidence-informed themed beliefs and perception categories relevant to the prevention of work disability. These themes and their working definitions are shown in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. The three evidence-informed beliefs and perceptions themes and their original working definitions.

https://doi.org/10.1371/journal.pone.0327355.t001

Some items from the initial pool of 22 items were deemed to fit within both themes A and B, such as ‘I believe that my condition will interfere with my ability to work’. In these instances, the item was allocated to the theme that it seemed to best fit conceptually. In addition, some redundancy was recognised which resulted in some of the items being excluded.

The final number of items was 12 across 3 scales, including 6 items for Health Beliefs (Category A), 3 items for Work Beliefs (Category B), and 4 for Employer Perceptions (Category C). One of the items was deemed relevant for both the Health and Work Beliefs dimensions (“It is not really safe for me to work”) and included as such in the scoring of both dimensions. An additional single item measure assessing work-related Recovery Expectations (“I am confident that I will be working/ will still be working in 3 months”) was also included in the Positivum: Beliefs and Perceptions measure. While this item was relevant to a Beliefs and Perceptions measure in this context, it stood on its own conceptually in that it did not seem to be a good fit for any one of the three beliefs and perceptions theme/ categories; past research supports that this single item is predictive of poor outcomes for those with pain-related injuries [14].

Administration and scoring

The questionnaire was designed for multiple modes of administration. Face to face administration was preferred, with an interviewer-administration over the phone or self-administration using a web-based data collection form also available to those receiving OR services in regional or remote areas.

Each item is scored on a 5-point Likert scale where 1 = strongly agree and 5 = strongly disagree. Positively expressed items are inverse-scored, such that a higher score always reflects a more positive response.

Participants and procedure

For the purposes of this study, de-identified data from a total of 3,352 individuals with musculoskeletal injuries were extracted for research purposes on 24^th May 2024 from an existing database containing information routinely collected as part of occupational rehabilitation and return to work service provision through MedHealth. Following referral by compensation and insurance regulators to MedHealth, clients routinely complete the Positivum assessment tool, including the PBPS, as part of a comprehensive initial assessment. A link is emailed to the client who completes the survey (typically during the initial face-to-face assessment in the presence of the OR consultant), the results subsequently made available for review during the next meeting with the trained allied health consultant to inform service delivery.

This study was approved by MedHealth Executive and consultation sought regarding privacy and consent from legal counsel. Individual written, signed consent is routinely obtained from clients at the beginning of service provision and before any data collection as part of best practice, usual care as per standard procedures. Completed consent forms are uploaded to and retained on a secure customer records management system. Subsequent data analyses were conducted internally on de-identified data for the purposes of refining an existing measurement tool and associated OR service provision.

For the purposes of the current study, eligibility criteria included:

• Injured individuals receiving OR services through either workers’ compensation (WC) scheme or through motor vehicle accident compulsory third party (CTP) insurance scheme
• Referral between 1st January 2020–30th April 2024
• Musculoskeletal condition as the primary injury or condition for receiving OR services
• Aged between 18 and 64 years

Note also that if an eligible individual provided responses to the Beliefs and Perceptions questions on multiple assessment occasions, only the earliest assessment was included in the study.

Statistical analyses

Only 12 items measuring the three themes of Beliefs and Perceptions were included in analyses reported here. Missing data for the Beliefs and Perceptions items were minimal, with four items (10, 11, 12, and 13) each missing one response. All available data were included in analyses. Unless otherwise indicated, results of statistical tests were interpreted at 0.05 alpha level.

Exploratory factor analysis

EFA aimed to examine the number and content of distinct dimensions present among the 12 Beliefs and Perceptions items. The measure was hypothesised to represent three dimensions, but there was uncertainty about separation of Work Beliefs and Health Beliefs dimensions, with one item initially allocated to both dimensions. The goal of EFA is to find the smallest number of underlying dimensions (factors) that account for the relationships between items, with the amount of information represented by each factor termed its eigenvalue. Three tests were used to find an optimal number of factors: number of factors with eigenvalues >1.0 (a factor carries more information than any individual item) [15], visual examination of scree plot to identify an abrupt change in the magnitude of factor eigenvalues (retaining only factors above the break) [16], and parallel analysis [17]. Parallel analysis involves simulating multiple sets of random data (200 in the present study), which have the same number of variables and participants as real data, but no underlying factors. Each random dataset is factor-analysed and the resulting eigenvalues are averaged and compared with real data. Eigenvalues from real data that exceeded those from the random data are used to indicate the number of factors to retain.

Where dimensionality criteria disagreed, all competing solutions were assessed on: (1) simple structure – each item has a substantive (>0.32, in absolute value) [18] loading on only one factor, (2) conceptually meaningful factors, and (3) adequate variance explained by the solution overall (>50%). Communalities, which capture the amount of variance explained by a given factor solution in individual items, were also considered, with values >0.2–0.50 interpreted as good and >0.50 excellent. Items that had no substantive loadings on any of the extracted factors, cross-loading items (loading on more than one factor), and items with communalities <0.20 were considered for removal, pending results of item analyses and consideration of whether their removal jeopardised content validity of measure.

Factors were extracted with Robust Weighted Least Squares (WLSMV) estimator [19] for categorial data. Crawford-Ferguson (C-F) oblique Varimax rotation [20] was used to aid factor interpretation. To assess sensitivity of the derived solution to the choice of extraction/rotation combination, analyses were repeated with restricted maximum likelihood (MLR) extraction and C-F Geomin rotation. In interpretation of each EFA solution, factor loadings adjusted for correlations between factors (pattern coefficients) were utilised.

Item analysis

IRT analyses aimed to assess measurement properties of items within the scales identified with EFA. In the case of a theoretically multifactorial item (4 “It is not really safe for me to work”), the item would be allocated to a scale for which it had the highest loading in EFA. The Rasch [21] model was applied to summarise the relationship between the level of a trait represented by an item (item difficulty, sum of responses to a given item across all persons) and the level of a trait expressed by a person (person ability, an individual’s score across all items). The specific type of the Rasch model applied in our study was partial credit model [22]. The original form of the Rasch model [21] was developed for dichotomous items, with various extensions subsequently proposed for items with more than two response categories. Of these, the partial credit model and the rating scale model [23] have been specifically developed for Likert-type items. A major difference between the two models is their handling of category threshold locations. Thresholds represent points on the ability continuum where two adjacent response categories are equally likely to be endorsed. Under the rating scale model, all items are constrained to have the same threshold locations whereas the partial credit model allows threshold locations to vary between items. While the rating scale model tends to be computationally less intense due to fewer parameters to be estimated (i.e., only one set of thresholds is estimated across all items), constraining thresholds to be the same across all items can potentially mask the presence of items with category disordering, where a higher response category is associated with a lower level of underlying ability. On the other hand, the partial credit model allows category thresholds to be examined for each item individually. Since category ordering was of major interest to this study, we utilised partial credit model in our IRT analyses, despite it being computationally more intense than the rating scale model.

Under the Rasch partial credit model, items and persons are positioned along a continuum, with the more able individuals expected to consistently endorse higher response options on the more difficult items. In our study, lower overall scores for items (indicating that item is more difficult to disagree with) and persons captured the less positive beliefs and perceptions about work, while higher item and person scores represented the more positive end of beliefs and perceptions.

IRT analyses assessed unidimensionality of emergent scales, their targeting ability, category ordering, and overall fit of items with their intended scales. Differential item functioning (DIF) analyses were also undertaken to examine possible presence of measurement bias. Items displaying category disordering, misfit to the Rasch model, and evidence of measurement bias were earmarked for removal.

Unidimensionality of each scale was assessed through the principal component analysis of residuals (PCAR). PCAR involves factor-analysing item correlations that are unaccounted by the primary dimension. Unidimensionality is supported when the amount of variance explained by the first residual factor is small relative to the amount of variance explained by the primary factor. Since there is no agreed definition for an acceptable amount of variance in the residual factor, we compared the results of PCAR from the observed data with those derived from data simulated to fit the Rasch model. For each scale, five simulated datasets were created, with values averaged across datasets. Scales were considered unidimensional if 1) primary dimension explained >50% of variance; 2) eigenvalue of the first residual factor was < 2.0; 3) the residual factor explained approximately the same amount of variance as that from simulated data. Overall goodness of fit of a scale with the Rasch model was consistent with non-significant log-likelihood (LL) chi-square, normed chi-square (chi-square/degrees of freedom) values 2.0–5.0, and person separation index (PSI) ≥2.0, which would indicate that the scale is able to differentiate between people with high and low levels of a trait [24].

Targeting represents how well scale items capture the full continuum of abilities in the population of interest. When the difficulty continuum represented by the items does not match the continuum of respondents’ ability, the underlying trait is measured imprecisely, and content validity of a scale may be jeopardised. Targeting was assessed by visually examining the distribution of person abilities and item difficulties and by comparing means (M) and standard deviations (SD) of ability levels of respondents with those of scale items.

Category ordering was assessed to examine functioning of response categories. When response scale is functioning as intended, response categories are hierarchically ordered from lowest to highest, with each category displaying a distinct peak on trait continuum. Category disordering is inferred when higher response categories are subsumed under the lower categories and may indicate that respondents have difficulty discriminating between the response options.

Item fit was assessed with mean square (MS) and z-statistics. Infit (information-weighted) and outfit (outlier-sensitive) fit statistics were utilised. MS values between 0.5 to 1.5 and z-statistics −2.5 to 2.5 were considered optimal [24]. Values below this range are indicative of possible redundancies and high values characterise “noisy” items (high measurement error). High measurement error is detrimental to quality of measurement and noisy items were considered for removal. Redundancies occur when the level of a trait targeted by an item overlaps with one or more other items. While redundancies do not degrade quality of measurement, the presence of several redundant items can make a scale unnecessarily long. In the presence of item redundancies, residual correlations >0.30 between items were used to flag pairs of closely related items, with a misfitting item of the pair considered for removal.

DIF occurs when groups defined by characteristics such as age or gender, for example, have different scores on an item after controlling for the overall score. The presence of DIF was assessed across groups defined by age (<45 vs ≥ 45 years), sex, body part affected (back vs other; lower limb vs other; upper limb vs other); work status (currently at work vs no or unknown), and socio-economic status (bottom 50% of postcodes on the Index of Relative Socio-economic Advantage and Disadvantage vs top 50%) [25]. A combination of DIF contrast ≥0.50 logits and significant Mantel-Haenszel chi-square was interpreted as evidence of measurement bias. To reduce the probability of false positives, Bonferroni corrected alpha (0.05/number of items) was used to interpret the significance of chi-square tests.

Internal consistency

Following EFA and item analysis, Cronbach’s α [26] and McDonald’s ω [27] were calculated for each emergent scale. Cronbach’s α represents the “true score” portion of the observed score, and despite being one of the most frequently utilised measures of internal consistency, has been shown to produced biased estimates of true reliability of a measure under certain conditions [28,29]. On the other hand, McDonald’s ω, which is derived from CFA factor loadings, tends to give a closer approximation of true reliability of a scale than α under a range of conditions. For both coefficients, values above 0.70 were regarded as acceptable [30].

Both coefficients are reported with 95% bootstrap confidence intervals (95% CI), derived from 1000 bootstrap draws. Factor loadings and item residual variances for the calculation of ω were derived from single factor CFA models for each scale, with no residual correlations between items allowed as future use of scale scores is not expected to take into account item residual correlations.

Confirmatory factor analysis

The goals of CFA were two-fold: to assess factorial validity of Beliefs and Perceptions measure and to evaluate measurement equivalence of the scales between individuals receiving musculoskeletal OR services through a workers’ compensation scheme and those receiving the services through the motor vehicle accident insurance scheme (WC and CTP groups, respectively). The number and content of factors for CFA was informed by EFA and IRT analyses. For model identification, factor variances were made equal to those of indicator items. In interpretation of CFA models, fully standardised coefficients were used. Factorial validity of Beliefs and Perceptions measure was supported when the hypothesised CFA model showed good overall fit with the observed data. Good fit was consistent with a non-significant chi-square test, normed chi-square 2.0–5.0, comparative fit index (CFI) and Tucker-Lewis index (TLI) >0.95, and standardised root mean square residual (SRMR) <0.08 [31,32]. Additionally, root mean square error of approximation (RMSEA) values <0.06, 0.06–0.08, 0.08–0.10, and >0.10 were interpreted as good, reasonable, mediocre, and poor fit [32,33], respectively.

Fit of CFA models was also evaluated at an item level with misfit indicated by low (≤0.32) or non-significant loading on the hypothesised factor, high uniqueness (<50% of variance shared with the rest of scale items), several residual correlations >0.05 with other items, and several modification indices (MI) >10. Misfitting items were considered for removal, if conceptually appropriate.

Testing of CFA models proceeded in a hierarchy. Single-factor model(s) were tested first in each study group, followed by a multi-factor model.

In the final step of analyses, measurement equivalence of Beliefs and Perceptions scales was evaluated in multi-group CFA to determine whether scales functioned in the same way in CTP and WC populations. Measurement equivalence testing followed a step-up approach, beginning with a configural model (least constrained model; number and composition of factors equal between groups, but factor loadings and item thresholds allowed to vary), metric model (factor loadings are equal between groups), and scalar model (factor loadings and item thresholds equal between groups). Fit of progressively more restrictive models was compared using chi-square difference tests, with significant results indicating that the scales did not function in the same way in the two groups.

While the goal of CFA was to test fit of Beliefs and Perceptions scales, rather than to find a well-fitting CFA model, severe misfit at the level of single-factor models was undesirable, as it would also affect the fit of subsequent multi-factor and measurement equivalence models. This is due to the more complex models presenting more opportunities for the model misfit to arise. Therefore, in the presence of suboptimal fit of single-factor models and prior to testing more complex CFA models, model modifications in the form of within-factor correlated residuals were introduced (no more than one per factor) if these made sense conceptually and improved model fit. However, no cross-factor loadings or cross-factor correlated residuals were permitted.

Sample size

IRT analyses generally require sample size of 100 + individuals [34]. EFA and CFA are considered large sample techniques, although there are currently no agreed criteria for optimal sample size. Generally, larger sample sizes are required for EFA when the underlying dimensions are not strongly distinct, or when the number of items defining factors is not high [35]. A previous Monte Carlo simulation study suggested minimal sample size of 200 to be adequate for the purposes of fitting a theoretical CFA model to the data [36]. In the present study, there was reasonable doubt about conceptual distinctness of Work Beliefs and Health Beliefs dimensions, due to a common item. Based on considerations outlined in [35] and results reported in [36] optimal sample size for the study was estimated to be 400 individuals for both EFA and CFA in each subpopulation, with minimal acceptable sample size of 200 per subpopulation [35,36]. Since conducting EFA and CFA on the same sample could adversely impact reproducibility of factorial structure of a measure in future studies, these analyses were carried out on different samples of individuals. As a result, there were four study samples in total: calibration WC, calibration CTP, holdout WC, and holdout CTP. EFA and item analysis were carried out with calibration samples while internal consistency and CFA utilised holdout samples.

Statistical software

EFA and CFA were carried out in Mplus version 8.5 [37] and item analyses were carried out in Winsteps version 5.8.0 [24]. Parallel analysis was carried out with Monte Carlo PCA for Parallel Analysis [38]. Remaining analyses were carried out in Stata 18 [39].

Participants

Characteristics of the study population are summarised in Table 2 and descriptive statistics for the Positivum Beliefs and Perceptions items are in Table 3. The database contained responses of 3,352 eligible individuals. Of these 2,978 were receiving OR services through the WC scheme and 374 were part of the CTP insurance scheme. All CTP clients were included in the study, with 174 randomly allocated to a calibration sample and the remaining 200 allocated to a holdout sample, as a larger sample was deemed more important for CFA. Of the WC clients, 400 each were randomly selected for calibration and holdout samples, respectively (S1 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Demographic and clinical characteristics of the study population.

https://doi.org/10.1371/journal.pone.0327355.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Summary of responses on beliefs and perceptions items.

https://doi.org/10.1371/journal.pone.0327355.t003

Median age of the study population was 44 years (interquartile range 32–54 years) and slightly more than half were male (55.6%). The CTP group were slightly younger (median age 41 years), less likely to be in paid employment (12.3% vs 29.8%), and had higher proportion of multiple injuries (65.2% vs 9.3%) than the WC group.

Exploratory factor analysis

In both WC and CTP samples, three factors had eigenvalues > 1.0. Scree plots also suggested three factors. However, only two factors in each study sample had eigenvalues greater than those of random factors (S2 Table and S1 Fig).

In three-factor solution (S3 Table), items measuring Employer Perceptions emerged as a distinct factor (Factor 3). Factor 2 was represented by two Health Beliefs items (item 9 “My condition gets in the way of me doing things I want to” and item 10 “I believe that my condition interferes with my ability to work”) and explained only 10% of variance in the data. Factor 1 was represented by the remaining Health Beliefs and Work Beliefs items. Item 13 (“There are things I enjoy about working/ think I would enjoy about working”) was cross-loading on factors 1 (0.40 WC, 0.49 CTP) and 2 (0.34 WC, 0.45 CTP).

In two-factor solution (Table 4), factor 1 captured Health Beliefs and Work Beliefs items and factor 2 contained Employer Perceptions items in both samples. Also in both samples, item 13 had no substantive loadings on either factor. Three items – 9 and 10 in both samples and 12 in WC sample – were cross-loading.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Results of two-factor exploratory factor analysis solution for Positivum Beliefs and Perceptions (pattern coefficients).

https://doi.org/10.1371/journal.pone.0327355.t004

While neither two- nor three-factor solutions were ideal, following a team discussion, two-factor solution was deemed preferred due to its greater parsimony. The two factors were labelled Employer Perceptions and Health-related Work Beliefs. The solution explained 62.5% and 63.7% of variance in the WC and CTP samples, respectively. Apart from item 13, communalities were acceptable in both samples.

Comparable results were obtained with MLR extraction/Geomin rotation. However, three-factor solution produced a Heywood case, with an item loading >1.0 in the CTP sample (item 7), suggesting the extraction of too many factors, thus further supporting two-factor solution as being more robust mathematically.

Item analysis

For item analysis, two scales were specified and analysed separately: Employer Perceptions (consistent with the initial theme) and Health-related Work Beliefs (Work Beliefs and Health Beliefs items combined). Although item 13 had no loadings on either of the two factors and item 10 had a slightly larger loading on Employer Perceptions factor in CTP sample, these items were analysed within the Health-related Work Beliefs scale due to a better conceptual fit.

Employer perceptions scale

Unidimensionality and overall fit.

PCAR supported unidimensionality of Employer Perceptions scale in each study sample. Variance explained by the primary dimension was adequate (WC: 57.7%, eigenvalue = 5.3; CTP: 59.2%, eigenvalue 5.8). The first residual factor had eigenvalue 1.6 in both samples and this was comparable with simulated data (WC: 1.4; CTP:1.5). Percentage of variance explained by the residual factor (WC: 17.0%; CTP: 16.0%) was slightly higher than for simulated data (WC: 11.4%; CTP: 11.8%) and chi-square test was significant (both p < 0.001) indicating misfit with the Rasch model in both samples (Table 5). However, normed chi-square values were within the acceptable range (WC: 2.0; CTP: 1.9), suggesting that the misfit could at least partially be attributed to the relatively large sample sizes. PSI values were 2.2 in WC sample and 2.3 in CTP, indicating that the scale can reliably differentiate between individuals with high and low levels of employer perceptions positivity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Item fit statistics from item response theory item analyses.

https://doi.org/10.1371/journal.pone.0327355.t005

Targeting.

Mean person and item locations were reasonably close in both samples (WC: persons M = −0.1 SD = 2.1, items M = 0.0 SD = 0.3; CTP: persons M = −0.7 SD = 2.2, items M = 0.0 SD = 0.3). In both samples, levels of employer positivity represented by items were somewhat narrower than trait levels of individuals (Fig 1), indicating less measurement precision at high and low levels of employer perceptions positivity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Joined distributions of person abilities and item difficulties for Beliefs and Perceptions measure.

Abbreviations: C/WC = calibration sample, Workers Compensation scheme; C/CTP = calibration sample, Compulsory Third Party insurance scheme.

https://doi.org/10.1371/journal.pone.0327355.g001

Category ordering.

Positioning of response categories along the continuum of employer perceptions positivity is shown in Fig 2 (also see S2 Fig). In both samples, response categories for each item displayed a distinct peak along the trait continuum, with no evidence of category disordering.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Ordering of response categories along a continuum of Beliefs and Perceptions levels.

Note: Items with category disordering are highlighted. Abbreviations: C/WC = calibration sample, Workers Compensation scheme; C/CTP = calibration sample, Compulsory Third Party insurance scheme.

https://doi.org/10.1371/journal.pone.0327355.g002

Health-related work beliefs scale

Item removal

No items were removed from the Employer Perceptions scale. While item 5 showed evidence of misfit on standardised measures of fit, the misfit was mild and was only present in the larger of our two samples (WC sample). Since z statistics are highly influenced by sample size, we regarded the misfit to not be detrimental to measurement properties of the scale. Additionally, item 5 had the highest difficulty score of all items in the Employer Perceptions scale and the removal of this item would further decrease the targeting ability of the scale, impacting its content validity. Hence, item 5 was retained.

In Health-related Work Beliefs scale, item 10 had high amount of measurement error and showed evidence of category disordering. As its content overlapped with that of item 9, item 10 was removed from the Beliefs and Perceptions measure. Item 13 also showed poor fit, but it was nonetheless considered important to Beliefs and Perceptions measurement as it was the only item that captured perceived enjoyment of work. It was therefore decided to leave item 13 as a single-item measure. Removal of items 10 and 13 did not impact the targeting ability of Health-related Work Beliefs scale, with PSI of 2.9 and 3.2 in WC and CTP samples, respectively. Table 6 presents the revised 12 item Positivum Beliefs and Perceptions Scales with refined definitions, and S4 Table within the supplementary information presents the revised Beliefs and Perceptions items as they pertain to the revised scales.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Revised 12 item Positivum Beliefs and Perceptions Scales (PBPS) with definitions.

https://doi.org/10.1371/journal.pone.0327355.t006

Internal consistency

Both Employer Perceptions and Health-related Work Beliefs scales had excellent internal consistency. Cronbach’s α for the four-item Employer Perceptions scale was 0.80 (95% CI 0.77–0.84) in WC sample (n = 400) and 0.83 (95%CI 0.79–0.88) in CTP sample (n = 200). Cronbach’s α for the six-item Health-related Work Beliefs scale was 0.90 in WC (95% CI 0.89–0.92) sample and 0.89 (95%CI 0.86–0.91) in CTP sample. McDonald’s ω for the Employer Perceptions scale was 0.84 (95% CI 0.82–0.85) and 0.87 (95% CI 0.86–0.88) in the WC and CTP samples, respectively. For the Health-related Work Beliefs scale, McDonald’s ω was 0.93 (95% CI 0.93–0.94) and 0.92 (95% CI 0.91–0.93) in the WC and CTP samples, respectively.

Confirmatory factor analysis

Single-factor models.

Fit of CFA models tested in this study is summarised in Table 7. Single-factor models for Employer Perceptions scale showed a moderately good fit in both samples. CFI, TLI, and SRMR were within the acceptable range, but chi-square tests and RMSEA indices indicated misfit in both samples (Table 7, Model 1.1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Fit indices for the confirmatory factor analysis models tested in the study.

https://doi.org/10.1371/journal.pone.0327355.t007

While the fit of Employer Perceptions scale was adequate overall, an improvement in the overall fit of a single-factor CFA model could potentially facilitate the interpretation of fit indices for the subsequent models of greater complexity, including a multi-factor CFA model and measurement equivalence testing. Hence, residual correlations and MI were inspected to identify possible model modifications that could potentially improve the fit of the current single-factor model prior to subsequent analyses.

There were no residual correlations above 0.05 in either sample, but MI indicated that correlations between residuals of items 5 and 8 or items 5 and 6 could potentially improve fit. Correlation of items 5 (“Employers prefer not to hire people with disabilities”) and 6 (“Because of my health, employers think that I am too much trouble”) was deemed conceptually defensible, as both items captured concerns with employer preferences for hiring people with disabilities. The modified model had a significantly better fit than the initial model in both samples (Table 7, Model 1.2).

Heath-related Work Beliefs scale also showed adequate overall fit to a single-factor CFA model, albeit it was also misfitting on chi-square and RMSEA tests in both samples, with a slightly better fit in WC sample (Table 7, Model 2.1). Similarly to Employer Perceptions scale, a slightly better fit could potentially facilitate the interpterion of fit indices of the more complex subsequent models. Inspection of residual correlations and MI suggested that model fit could be improved with a correlation between residuals of items 3 and 7. As both items expressed concerns about worsening health while working, the correlation was incorporated into the model, with a significant improvement in model fit (Table 7, Model 2.2), although chi-square and RMSEA remained suboptimal in CTP sample.

Two-factor model.

Two-factor model of Beliefs and Perceptions measure had adequate fit in both samples, supporting its factorial validity (Table 7, Model 3). While some misfit was still present on chi-square and RMSEA indices, high residual correlations and high MI were only present for cross-factor correlated errors and cross-factor loadings. Hence, no further model modifications were deemed appropriate. Path diagram of two-factor model for the combined WC and CTP holdout samples is shown in Fig 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Path diagram of the two-factor model of Beliefs and Perceptions measure from the combined holdout sample.

Model fit summary: χ²(32) = 195.8, p<0.001; normed χ² = 6.1; RMSEA = 0.09 (90%CI: 0.08-0.11; Probability RMSEA ≤0.05: p<0.001); CFI = 0.99, TLI = 0.98; SRMR = 0.03. Note: N = 600.

https://doi.org/10.1371/journal.pone.0327355.g003