A systematic review and a qualitative study were initially conducted to identify relevant factors used to construct a questionnaire for predicting TB patient non-adherence. Subsequently, a framework that described the five factors influencing long-term medication adherence, namely, socioeconomic status, healthcare team, medical conditions, therapy, and patients, was adopted from the WHO (World Health Organization, 2003).

Building upon the previously established framework, questionnaire items representing each dimension or variable to be measured were developed. Although the framework primarily pertained to adherence, this study adapted all the obtained dimensions to the context of non-adherence. Five items created for each indicator in the variable were assessed by TB treatment experts and analyzed by psychologists.

A panel of experts, including a psychologist, three pulmonary specialists, two nurses, and three pharmacists, assessed the level of difficulty and adequacy of the questionnaire. This evaluation was based on qualitative study activities conducted before the questionnaire development. The experts participated in focus group discussions (FGD) regarding the factors influencing non-adherence and were selected according to the possession of at least 1 year of experience in TB medication. Specifically, psychologists were engaged in assessing the readability and comprehensibility of the items before the pretesting stage. Following this process, the initial selection of items was refined based on assigned the median total score. Items that received higher scores, as indicated by multiple experts, proceeded to the next stage.

After the panel of experts conducted a content validity assessment, a pilot questionnaire was pretested on 50 respondents in December 2021, as recommended by healthcare professionals. Criteria for participant selection included a history of medication non-adherence, age 18 years or older, a minimum of high school education, and willingness to provide informed consent. Respondents were recruited during hospital visits, and each completed a paper copy of the questionnaire. Trained assistants reviewed the self-administered questionnaires on-site before being delivered to the study team.

Confirmatory factor analysis (CFA) is an integral component of SEM, valuable for the appropriateness of variable measurements concerning the number of factors. In CFA, factors can be considered as constructs, and this analysis represents an interdependence technique for determining the underlying structure in construct variables. High partial correlation in factor analysis holds practical and statistical significance, with the general rule of thumb suggesting values above 0.70 as conceptually valid (Lance and Vandenberg, 2002; Harrington, 2009; Brown and Moore, 2012). However, the Bartlett roundness test at a level of >0.05 indicates a sufficient correlation between construct variables for a single-factor analysis (Suhr, 2006; Hair et al., 2014a; Gatignon and Gatignon, 2014; Brown, 2015).

SEM is a statistical model that describes the relationships among several variables (Hair et al., 2014b). During the calculation process, SEM simultaneously examines the structural relationships expressed through a series of equations resembling multiple regression equations. These equations elucidate all the interconnections between analyzed constructs, comprising both dependent and independent variables. Constructs are unobservable and cannot be represented by numerous variables compared to those representing factors in CFA. Moreover, PLS-SEM is a causal-predictive method of SEM that stresses prediction in estimating statistical models aimed at providing causal explanations (Hair et al., 2019). In this study, each item used an ordinal Likert scale for measurement, with five potential response levels. Indicators with ordinal responses from at least four categories may be interpreted as intervals, or at the very least, as continuous variables. No two indicators for a construct must have the same scale type, and scale values need not be normalized (Hair et al., 2014b). The utility of a questionnaire as a study instrument is evaluated using the validity and reliability method. Validity refers to the extent to which observations accurately record the examined variables. Meanwhile, reliability relates to the consistency of observations, often determined by whether two (or more) observers or the same observers, monitoring the same event on successive occasions, reach similar conclusions (Sekaran and Bougie, 2016).

The five-dimensional framework of WHO for adherence was adapted into the concept of non-adherence in this study. These five dimensions comprised the various causes and risk factors associated with non-adherence, based on the results of systematic reviews and qualitative studies conducted previously. Each dimension had specific indicators used for measuring its impact on medication non-adherence. A total of 32 indicators were successfully generated, with each contributing five items. As shown in Figure 2, the dimensions were as follows: socioeconomic (6 indicators; 30 items), healthcare team (8 indicators; 40 items), conditions (4 indicators; 20 items), therapy (9 indicators; 45 items), and patients (5 indicators; 25 items). This resulted in the generation of a total of 160 items in the subsequent stage.

The items generated for the questionnaire were adapted to local settings and divided into six main sections, as presented in Table 1. These included Demographic factors (5 items) which examined demographic and socioeconomic characteristics. Socioeconomic-related factors (30 items) constituting considerations such as economic priorities. In developing countries, patients with low socioeconomic status often face the challenge of balancing competing priorities. These competing priorities might require allocating limited resources to meet the needs of other family members, such as the children or parents catered for (Killewo, 2002; World Health Organization, 2003; Diniawati and Wibowo, 2018; Mahara et al., 2018). Healthcare team-related factors (40 items) explored the effects of the patient-provider relationship, and more investigations are needed concerning the impact of the healthcare team and system-related factors on non-adherence. While an excellent patient-provider relationship can increase adherence several factors have a negative effect. These are comprised of underdeveloped healthcare services, inadequate or nonexistent reimbursement by health insurance plans, poor drug distribution systems, and a lack of knowledge and training among healthcare providers in managing chronic diseases (World Health Organization, 2003; Do Peterson et al., 2012; Gugssa Boru et al., 2017). Condition-related factors (20 items) included demands, symptoms, and disease-specific issues targeted by healthcare professionals. Several conditional factors, such as patient geography and health status, influenced their willingness to complete medication (World Health Organization, 2003; Shargie and Lindtjørn, 2007; Tadesse et al., 2013; Woimo et al., 2017; Ruru et al., 2018b). Therapy-related factors (45 items) consisted of the main barriers to adherence found in intervention studies, such as dosing frequency and side effects. Collaboration between pharmaceutical companies, health professionals, and researchers is essential to address this issue. Health systems play an essential role in minimizing the impact of side effects (World Health Organization, 2003; Do Peterson et al., 2012; Gugssa Boru et al., 2017; Heuvelings et al., 2017). Patient-related factors (25 items) examined the primary barriers to compliance as described in the reviewed literature, namely, a lack of information and self-management skills, difficulties with motivation and self-efficacy, and inadequate support for behavioral change (Morisky et al., 1990; Chambers et al., 2010; Gugssa Boru et al., 2017). These barriers are specifically relevant for interventions aimed at changing habits and lifestyles, as well as influencing drug use. The WHO recognizes the need to support patient self-management efforts, and many researchers are working to develop, enhance, and disseminate self-management guidelines (World Health Organization, 2003; Gough and Kaufman, 2011; Van Den Boogaard et al., 2012; Chowdhury et al., 2015; Sahile et al., 2018).

This section consisted of choices on a scale from one to five, with response categories ranging from “strongly disagree” to “strongly agree” for each question. Following the results of the item selection by experts, a total of 60 questionnaire items were used in the pilot study. At this stage, the expectation was to obtain a questionnaire containing a more streamlined set of items to facilitate measurement with fewer items during validation.

In this phase, the initial selection of 160 items was reduced to 60, based on a median total score of 2.0 for each. Items selected by more than two experts proceeded to the next stage.

The results of face validity obtained during the pilot study featuring 50 respondents reduced the number of questionnaire items from 60 to 22. The response rate was 100% (50/50 participants), with respondents requiring an average of 15 min to complete the questionnaire. The validity of each item is presented in the Pearson correlation column in Table 2. Considering the 50 respondents (N) and a significance level of 0.05, the minimum Pearson correlation value was 0.278. Therefore, 22 items exhibited Pearson correlation values exceeding 0.2732, denoted by * or ** in the Pearson correlation column of the output table. As a result, 38 items were considered invalid, while 22 were validated.

The 22 questionnaire items identified during the face validity assessment were administered to a total of 346 sensitive TB patients as respondents. None of the patients from the pilot study were included in the validation phase. Table 3 presents an overview of selected items and themes that successfully passed face validity.

Overall, the results of the questionnaire development phase can be seen in Figure 3.

Table 4 presents the demographic characteristics of the 346 respondents who participated in this study. The most common age range found was 20–29 years (24.28%), with a mean of 39.71 (±10.71) years. Most of the respondents were male (54.91%). The educational background was predominantly high school (82.37%), while some (17.34%) completed tertiary education. A significant proportion was unemployed (39.30%) and a small percentage was health workers (0.76%). A substantial number came from private hospitals (38.15%), which was justifiable because these hospitals served as referral centers for pulmonary diseases in Jakarta.

Table 5 shows the results for each measured factor, as each item in all the factors was tested for factor loading to identify the influence of each item.

Standardized factor loadings were expected to exceed 0.70, but factor loadings in the 0.40–0.70 range could be evaluated. The evaluation conducted featured content validity by considering the impact exerted on the AVE gain and composite reliability. A higher factor loading value signified greater validity of the construct measurement (Hair et al., 2014b). Factor loadings must be statistically significant, with t values exceeding 1.96 for a 5% significance level (Hair et al., 2014b). To measure convergent validity, the AVE was used, with a threshold value of 0.50. A higher AVE value indicated more information obtained from the latent and reflected similarity in the latent construct (Hair et al., 2014b). Composite reliability and Cronbach’s alpha ought to have a value of at least 0.70, although a minimum of 0.60 is acceptable for exploratory studies (Hair et al., 2014b). Regarding cross-loading, each indicator should exhibit a stronger correlation with its construct than other constructs, indicating discriminant validity. This empirical standard ensures that a measured construct is distinct from other constructs (Hair et al., 2014b).

The results of the questionnaires at the validation stage determined the items that proceeded to the SEM-PLS modeling stage (Figure 4). Upon model simulation, differences emerged between valid items in factor loadings at the analysis stage and factor loadings on SEM. In the analysis stage, SEA23, SVA41, and SEA11 were deemed invalid. In the simulated SEM-PLS model, SEA23, SVA41, CAA13, CAA14, BNA34, CAA72, SEA11, and SEA15 were excluded. The three items, including SEA23, SVA41, and SEA11, remained invalid in both factor analysis and SEM despite sharing similarities. CAA13, CAA14, BNA34, CAA72, and SEA15 which were valid in the factor analysis became invalid in SEM. This showed that SEM examined the effect of each item on the factor measured, and the influence of the factor on non-adherence. Consequently, SEM yielded more invalid items compared to factor analysis.

In the simulated SEM-PLS model, only one factor, namely, condition, significantly influenced non-adherence. This indicated why all condition-related factor items remained entirely valid at the analysis stage. However, the therapy-related factors had no impact on non-adherence in the SEM-PLS model. In comparison to the other four, condition-related factors significantly influenced patient non-adherence to medication.

Following the SEM-PLS modeling process, 14 valid items remained across the five factors, including those related to socioeconomics (4 items; SCA21, SCA23, SCA24, CAA45), healthcare team (4 items; CAA41, CAA42, CAA43, CAA44), medical condition (3 items; CAA11, CAA12, CAA15), therapy (2 items; BNA31, CAA71), and patients (1 item; BRA21). Only condition-related factors significantly influenced non-adherence, and as indicated in Table 7, the constructs developed were reliable. The reliability test conducted in PLS applied Cronbach’s alpha and composite reliability techniques. Cronbach’s alpha measures the lower limit of the reliability value of a construct, while composite reliability estimates the actual value. Composite reliability is considered better at estimating the internal consistency of a construct. Moreover, the rule of thumb used for the composite reliability value indicated >0.7, and the obtained Cronbach’s alpha value exceeded 0.6 (Chin, 1998; Ghozali, 2016), signifying that all constructs had good reliability.

Overall, the results of the questionnaire validation phase can be seen in Figure 5.