Development and validation of the Health Promoting Behaviour for Bloating (HPB-Bloat) scale

Study design, sampling method, and participants

Using a cross-sectional study design with purposive sampling, we conducted our study between May 2018 and October 2019. A total of 520 people from the compound of the Hospital Universiti Sains Malaysia (HUSM), Kelantan, Malaysia were screened. Ultimately, 510 (98.1%) participants were deemed eligible and were recruited for the study. The participants consisted of patients, caregivers, accompanying persons, hospital staff, and students around HUSM. The inclusion criteria were: age 18 years and older, a functional bloating diagnosis (based on the Rome IV criteria or a clinical diagnosis), and/or experience of bloating at least once in the past 3 months (based on an answer to the verbal question “Have you ever experienced bloating?” and/or using a pictogram from the Rome foundation). A clinical diagnosis of bloating was made based on the experience of the physicians who managed functional GI disorders. The exclusion criteria were: an absence of a history of organic GI diseases (e.g., inflammatory bowel disease, GI infections, and colorectal cancer), a history of abdominal surgeries, taking drugs that may cause or worsen bloating (e.g., opiates), and any severe psychiatric illnesses (e.g., schizophrenia). Appropriate eligibility criteria is important and can impact external validity when designing a study (Patino & Ferreira, 2018). These specific criteria were chosen in order to cover the adult general population that were not affected by severe diseases.

Ethical approval

Ethical approval was obtained from the Human Research Ethics committee, Universiti Sains Malaysia (USMKK/PPP/JEPEM/17010012) prior to the start of the study. This study also conformed to the guidelines set by the International Declaration of Helsinki. Written informed consent was obtained from each participant.

Developing the HPB-Bloat

The new HPB-Bloat scale was developed to measure health-promoting behaviours across people with AB, and was based on the TPB, one of the most commonly-used and well-validated decision-making models that examines attitude and behaviour. We employed the theory, driven with the approach of dimension/indicator analysis that was described by Hox (1997). Based on our research of the literature and discussions with the research team experts, we conceptualized AB health behaviours across five domains: diet, health awareness, physical activity, stress management, and treatment. The new item generation was conducted by the researchers through an extensive literature review related to health behaviours that could encourage improvements in AB symptoms. Based on the literature review, a total of 24 items were generated. The research team experts also provided an additional 10 related items and supported the five temporary domains from the early draft of the HPB-Bloat scale. In order to cover all of the important indicators for the behaviour construct, we conducted an in-depth interview of 12 individuals with AB symptoms. The in-depth interview was conducted using guided questions. For example, “What do you think about AB in daily life?”, and “How do you manage AB in daily life?” Additional probing questions that focused on specific activities used to manage symptoms included “How about your dietary intake? Does it contribute to improve your AB symptoms?”, “How about physical activity? Does it contribute to improve your AB symptoms?”, “How about stress management? Does it contribute to improve your AB symptoms?”, and “Are there any other things that help you deal with AB? If so, is it helpful and how does it help?”. The duration of the interview was approximately 30 min to 1 h. All the recorded interviews were transcribed into a transcript, which was then narratively analysed. Themes were identified from the transcript, a theme list was created, and interview segments were coded. Important aspects and critical points from the interviewed individuals were identified. From these interviews, we found an additional item that we added to the HPB-Bloat’s item pool. Hence, a total of 35 items were generated in the initial stage of developing the first draft of the HPB-Bloat. The responses for each item were rated using a five-point Likert-scale, from never (1) to very often (5). All items were developed in the Malay language, which is the main spoken language in the study’s location. The first draft of the HPB-Bloat was then examined for its content validity by seven invited experts, who each had at least 10 years of experience in the GI field, psychometric testing, language, and questionnaire development. Figure 1 shows the item generation process from the initial stage of development to the final stage of item reduction for the newly developed HPB-Bloat.

Content validity and pre-testing of the HPB-Bloat

To assess content validity, we computed content validity indices for the items (I-CVI) and scale (S-CVI) based on the relevant responses from the invited experts (Lynn, 1986; Polit & Beck, 2006). Using the average method (S-CVI/Ave), we found that the items’ content validity index and the scale’s content validity index for the five expected domains (or factors) in the HPB-Bloat were all more than 0.75, which were considered satisfactory (Lynn, 1986). Subsequently, we conducted pre-testing on 30 participants diagnosed with functional bloating (based on Rome IV or a clinical diagnosis). The invited participants were asked to comment on the clarity and comprehensibility of the administered HPB-Bloat. The format and font sizes were modified based on suggestions from the participants. However, the structure and wording of items remained unchanged.

Based on the results from the CVIs (using I-CVI and S-CVI/Ave) and pre-testing, the first version of the HPB-Bloat remained at 35 items. In order to determine the validity and reliability of this HPB-Bloat draft, we performed an exploratory factor analysis (EFA), followed by a confirmatory factor analysis (CFA).

Data and statistical analysis

The EFA and CFA were performed using the Statistical Package for Social Sciences (SPSS) version 24.0 (IBM, Armonk, NY, USA) and Mplus 8 (Muthén & Muthén, 1998). The variables were expressed as mean and standard deviation (SD) for numerical items and frequency (n) and percentage (%) for categorical variables.

The EFA was first performed on the initial 35 items in order to explore and extract the major contributing factors, using the principal axis factoring with Promax rotation (Kappa 4). We decided the number of factors based on the eigenvalue, which should be greater than one. Factor loadings greater than 0.40 were regarded as significantly relevant and items with lower loadings were considered problematic and were deleted (Hair et al., 2010). All item deletions were conducted in a consecutive manner and subsequent models were re-specified following each deletion. The reliability of the factors was checked using internal consistency based on Cronbach’s alpha, with the generally recommended threshold value of 0.60 (Taber, 2018). However, a value above 0.50 is still considered acceptable by some literature (George & Paul Mallery, 2003; Hinton, Brownlow & McMurray, 2004; Karin, Sabine & Marcel, 2019).

Following EFA, we tested the final structure model with CFA to further confirm its validity. Items with factor loadings less than 0.40 were removed after adequate theoretical support was carried out by researchers. The modification indices (MI) were used as a guide to improve the model by adding the items’ residual correlation. Modifications to the model were based on theoretical basis and expert opinion.

The measurement CFA model was assessed using several fit indices (Kline, 2011). The fit indices and the recommended threshold values were: a standardized root mean squared residual (SRMR) value lower than 0.080, a root mean square error of approximation (RMSEA) value lower than 0.070, and a comparative fit index (CFI) and Tucker-Lewis Fit Index (TLI) values above 0.92 for number of items 12 to 30 (Hair et al., 2010). The CFA model was re-specified with adequate theoretical support until most of the model fit indices met the criteria.

We used the correlation matrices among the latent constructs to establish discriminant validity. The composite reliability of the measurement CFA model was computed (Hair et al., 2010). Composite reliability greater than the threshold value of 0.60 was considered acceptable (Hair et al., 2010).

The EFA model

Our results indicated adequate sampling and reliable estimates (Kaiser-Meyer-Olkin Measure of Sampling Adequacy (KMO) = 0.732, Bartlett’s test of sphericity= p < 0.001; Kaiser, 1974; Yong & Pearce, 2013). The 35 initial items, divided into five factors, were represented with a cumulative value of 56.7% of the variance, which indicated acceptable importance. The variance value for each of the five factors was 22.25, 10.91, 9.58, 7.26, and 6.65, respectively.

We removed the problematic items from the subsequent EFA models until we achieved a final model with all factor loadings greater than 0.40. Three items with factor loadings less than 0.40 (T5 = 0.349, SM1 = 0.287, and D2 = 0.370) were kept for further analysis. Eventually, we removed 15 items and the final model consisted of 20 items. The descriptive statistics, factor loading, and internal consistency results of the final EFA model are summarized in Table 1. The five factors described were diet, health awareness, physical activity, stress management, and treatment. The internal consistency corresponded with the reported Cronbach’s alphas, ranging from a low of 0.52 (stress management) to a high of 0.81 (health awareness).

The CFA model

As shown in Table 2, the results of the initial 20-item CFA model revealed that none of the model fit indices met the criteria. Additionally, the CFA output showed that the latent variable covariance matrix was not positive definite, and there was a problem involving the treatment factor. Therefore, we subjected the initial model to re-specification for improvement. The CFA output indicated that the treatment factor had a high standardized correlation (>1) with the diet factor. However, we found that both factors had items that made them not suitable to be combined into one factor. We further investigated the meaning of the four items under the treatment factor. We removed these items iteratively from the initial CFA model, except for item T4 (“I will always be ready to try new treatment techniques to improve AB symptoms when needed”), which has a treatment awareness component. Therefore, we grouped item T4 under the health awareness factor and renamed the factor as health and treatment awareness. After removing the treatment factor, Model-1 consisted of four factors with 17 items (Table 3). The fit indices were improved in Model-2, but CFI and TLI were still not within the acceptable threshold values. Based on the MI results and following adequate theoretical support carried out by the researchers, we added residual correlations from several items within the same factor into the model, one by one. The final model (Model 2) fit the data well based on several fit indices: CFI = 0.929, TLI = 0.911, SRMR = 0.052, and RMSEA = 0.044 (0.032, 0.061) (Table 2, Model-2).

As Table 3 illustrates, all standardized factor loadings exceeded the threshold of 0.40. The composite reliability of all Model-2 factors was greater than 0.60, which indicated good reliability.

As shown in Table 4, all correlations between factors were below 0.85, which suggested that the HPB-Bloat’s discriminant validity was satisfactory.