The reliability, validity and factorial structure of the Swahili version of the 7-item generalized anxiety disorder scale (GAD-7) among adults living with HIV from Kilifi, Kenya

This work is part of data collected from a larger cross-sectional study looking at different outcomes in PLWHA, including health-related quality of life and mental health. The cross-sectional study was conducted at the Center for Geographic Medicine Research (CGMR), located in Kilifi County, between 20th February and 15th April 2018. The centre is within Kilifi County Hospital, the main referral hospital. In Kilifi County, prevalence of HIV in adults is estimated as 4.5% [26]. Participants were adults living with HIV attending an HIV specialized clinic at the Kilifi County Hospital. To be included in the study, they had to be 18 to 60 years old, with confirmed HIV-positive status, on antiretroviral therapy, and able to comprehend and/or communicate in the national language (Kiswahili) since this language was used during administration of all study instruments and asking for consent. We did not include elderly individuals (above 60 years) living with HIV because of the increased likelihood of illnesses associated with advancing age, which may have an impact on quality of life [27], an outcome of interest in the larger study. In total, 512 participants visiting the HIV clinic for their appointments were approached to participate, 44 of whom declined participation, whereas 18 were excluded because they were over 60 years of age (n = 11) or could not comprehend or communicate in Kiswahili (n = 7). The remaining 450 adults living with HIV participated in the study.

In line with international guidelines for translation of tools in health research (https://​www.​who.​int/​substance_​abuse/​research_​tools/​translation/​en/​), all questionnaires were independently translated from English to Swahili by two staff members fluent in both languages. Back-translations into English were then done by another independent pair of translators. To ensure content, conceptual, semantic and idiomatic equivalence of the questionnaires [33], a committee of HIV researchers (natives of Kenya, knowledgeable about the Kenyan culture, bilingual and fluent in both English and Kiswahili) and the translators then held a harmonization meeting. Discrepancies in the translations were resolved by consensus. The final versions were obtained following incorporation of changes resulting from pretesting procedures.

Cronbach’s alpha was used to evaluate internal consistency of Swahili GAD-7. ICC assessed test–retest reliability (using data of 60 participants rescreened 2 weeks following initial assessment). Using structural equation modelling, confirmatory factor analysis (CFA) with diagonally weighted least squares was used to determine the factor structure of Swahili GAD-7. An a priori model was specified, where one factor underlies the item responses and each item loads onto this unidimensional factor. The loadings associated with responses from the Swahili GAD-7 were then generated. The choice to confirm a one-factor structure was informed through an exploratory factor analysis (EFA) that checked for a possibility of the posited two-dimensional structure of GAD-7 [34, 35]. We examined the goodness-of-fit between the hypothesized model and the data using the following fit indices: root mean square error of approximation [RMSEA], comparative fit index [CFI] and Tucker–Lewis index [TLI]). Using multiple indices provides a conservative and reliable evaluation of model fitness [36]. Based upon the literature [37‐39], fit indices were considered acceptable if RMSEA ≤ 0.08, CFI and TLI ≥ 0.90, or excellent if RMSEA ≤ 0.06, CFI and TLI ≥ 0.95. Prior to CFA, we checked for sampling adequacy using Kaiser–Meyer–Olkin (KMO) value and Bartlett’s test of sphericity. We also examined inter-item correlations using Kendall tau-b correlation (in view of the ordinal measurement level of the items of GAD-7). Multi-group CFA was conducted to determine measurement invariance (configural, metric, and scalar invariance) between sexes (female versus male) and age groups (young adults 18–35 years versus middle-aged adults 36–60 years) for the hypothesized unidimensional factor structure of Swahili GAD-7. We used difference in CFI (ΔCFI), cut-off value ≤ 0.01, to define invariance as this is the most widely used and empirically best supported criterion [40‐42]. Convergent validity was assessed through correlation of GAD-7 scores with scores of both PHQ-9 and 12-item HIV stigma scale (Pearson’s product correlation). For discriminant validity, independent Student’s t-test followed by analysis of covariance (ANCOVA) were used to evaluate whether GAD-7 is sensitive to antiretroviral regimen (first-line versus second-line treatment). We used antiretroviral regimen as a proxy for perceived treatment success (switching to second-line treatment perceived as poor treatment success and expected to be related to higher GAD-7 scores) because of missing data on biological markers of disease progression (viral load or CD4 cell count) from the participant clinical records. All analyses were conducted using STATA (version 14.0) statistical software package except test–retest and CFA analyses which were done on R (version 3.4.1) software.

Table 1 summarizes sociodemographic, health and treatment characteristics of study participants by sex. Majority of the study participants were females (79.1%). The mean age of the 450 recruited participants was 42.7 years (Standard Deviation [SD] = 9.7). Participants’ median PHQ-9 score was 3 (Inter-Quartile Range [IQR]: 1—7). For perceived HIV-related stigma, mean score was 28.4 (SD = 7.7).

Table 2 presents the mean (SD) participant scores for each of the GAD-7 items alongside percentage response across the ordinal Likert scale options. Of important note is that responses were distributed across all the Likert options in the expected direction indicative of no major ceiling or floor effects.

Cronbach alpha for the Swahili version of GAD-7 revealed very good internal consistency, α = 0.82 (95% confidence interval [95% CI] 0.78, 0.85). Table 3 presents item–total correlations, the correlation of each item with the rest of the items, inter-item covariance and item alphas. Test–retest reliability of Swahili GAD-7 as indicated by intra-class correlation coefficient was acceptable, ICC = 0.70 (95% CI 0.55, 0.81).

Prior to CFA, we examined KMO and Bartlett’s test of sphericity for sampling adequacy and ran an EFA to check the possibility of a two-dimensional factorial structure. KMO value was 0.84 while Bartlett’s test of sphericity was highly significant χ² = 1057.32, p < 0.001. These results suggested that there were correlations in the dataset appropriate for factor analysis. Table 4 presents Kendall’s tau-b inter-item correlation coefficients. The correlations ranged from r = 0.3 to r = 0.7 and were all statistically significant at p < 0.001. EFA analysis showed a clear one-factor structure. The first factor was the only one with an eigen value above one (2.92) and explained 81.4% of the variance. Figure 1 shows the scree plot of the eigen values which suggests a one-factor structure best fit the data.

In CFA, all the GAD-7 Swahili items loaded well on this one-factor solution (item factor loading ≥ 0.5). Table 2 presents the factor loadings. Goodness-of-fit indices for the unidimensional factorial structure of Swahili GAD-7 were all excellent (RMSEA < 0.01 [95% CI 0.00, 0.05]; CFI = 1.00; TLI = 1.00).

Multi-group CFA was used to determine the extent to which the unidimensional factorial structure of Swahili GAD-7 was invariant across sex (female versus male) and age groups (young adults versus middle-aged adults). Overall, this multi-group analysis confirmed sex and age invariance of Swahili GAD-7.

In a model assuming the same item-factor assignment across sex and age groups (configural invariance), the one-factor solution of Swahili GAD-7 fitted the data well in both groups (Table 5). This is an indication that similar latent constructs have been measured in the groups hence the one-factor measurement model is acceptable for both sexes and age groups.

Assuming same item-factor assignment but also constraining factor loadings to equivalence across sex and age groups (metric invariance), all the goodness-of-fit indices for the one-factor solution of Swahili GAD-7 were excellent (Table 5). Comparing this metric invariance model to the least stringent configural invariance model, ΔCFI was < 0.001 and 0.009 for sex and age groups, respectively.

Assuming same item-factor assignment, but also constraining factor loadings and item intercepts to equivalence across sex and age groups (scalar invariance), all goodness-of-fit indices for the one-factor solution of Swahili GAD-7 were also excellent (Table 5). Comparing this scalar invariance model to the less restrictive metric invariance model, ΔCFI was < 0.001 and 0.008 for sex and age groups, respectively.

Evaluating all the goodness-of-fit indices and ΔCFI indicated that the scalar invariance model fitted the data best for sex group while metric invariance model fitted the data best for age group.

Scores of GAD-7, Swahili version, highly and moderately correlated with those of PHQ-9 (r = 0.73; p < 0.001) and 12-item HIV stigma scale (r = 0.36; p < 0.001), respectively. These patterns of results support convergent validity of Swahili GAD-7. Raw mean scores on Swahili version of GAD-7 were significantly higher among study participants on second-line antiretroviral treatment compared to those on first-line treatment (4.24 versus 2.60; p = 0.03). Using ANCOVA adjusted for age, sex and presence of any chronic medical illness, the observed difference remained statistically significant (F [1, 441] = 5.55, p = 0.02) suggestive of good discriminant validity of Swahili GAD-7.

Despite the study strengths in terms of sample size and methodological approaches to evaluating reliability, factorial structure and validity of Swahili GAD-7, it is not without limitations. We recruited participants from one primary care centre, thus these findings may not be generalizable to other settings or regions. We used a proxy indicator—antiretroviral regimen—to examine discriminant validity. Future studies involving PLWHA should confirm our finding by investigating for group differences on GAD-7 scores using either or both standard biomarkers of disease progression (viral load and/or CD4 cell count). We did not explore the diagnostic accuracy of Swahili GAD-7, therefore we recommend a separate study in this or similar setting, exploring the sensitivity and specificity of this measure at different cut-off values, using a diverse sample of PLWHA, to identify the optimal cut-off score for identifying those in need of mental health treatment or referral services in HIV clinical settings.