Religiosity and Mental Wellbeing Among Members of Majority and Minority Religions: Findings From Understanding Society: the UK Household Longitudinal Study

Abstract It is unclear whether links between religiosity and mental health are found in contexts outside the United States or are causal. We examined differences in mental wellbeing and associations between mental wellbeing and religiosity among the religiously unaffiliated, White and non-White Christians, Muslims of Pakistani, Bangladeshi, and other ethnicities, and other minority ethnoreligious groups. We used 4 waves of Understanding Society: the UK Household Longitudinal Study (2009–2013; n = 50,922). We adjusted for potential confounders (including socioeconomic factors and personality) and for household fixed effects to account for household-level unobserved confounding factors. Compared with those with no religious affiliation, Pakistani and Bangladeshi Muslims and members of other minority religions had worse wellbeing (as measured using the Shortened Warwick-Edinburgh Mental Wellbeing Scale and General Health Questionnaire). Higher subjective importance of religion was associated with lower wellbeing according to the General Health Questionnaire; associations were not found with the Shortened Warwick-Edinburgh Mental Wellbeing Scale. More frequent religious service attendance was associated with higher wellbeing; effect sizes were larger for those with religious affiliations. These associations were only partially attenuated by adjustment for potential confounding factors, including household fixed effects. Religious service attendance and/or its secular alternatives may have a role in improving population-wide mental wellbeing.

Web Appendix 1. Fixed-effect specification Equation for fixed effects analysis within households: ℎ = 0 + ℎ + ℎ + ℎ Where: • y is the outcome for each i individual within h household. • ℎ is a vector of household-variable exposures. • αh are household-level fixed effects that adjust for all time constant household level factors, such as socioeconomic factors. • ℎ is the individual-level residual term.
In Stata, the above can be estimated using xtreg, for example: xtreg outcome exposure [pweight = hholdweight], i(household_id) fe cluster(household_id) Web Appendix 2. Testing measurement invariance of SWEMWBS and GHQ across religions SWEMWBS Here, we test measurement invariance of the 7-item version of the Warwick-Edinburgh mental well-being scale across three religion categories: the non-religious, Christians of any denomination, and Muslims of any denomination. We exclude members of other religions from the analysis. We use the data from Understanding Society Wave 1. The seven items are (1) feeling optimistic about the future (2) feeling useful (3) feeling relaxed (4) dealing with problems well (5) thinking clearly (6) feeling close to others (7) able to make up own mind. A graphical representation of the measurement model is given below.
We fit four models with varying degrees of strictness of measurement invariance (see e.g. Kline 2016). The first one is the configual invariance model, which keeps the model structure in the above depiction the same, but allows the parameters to vary across the three groups. The second is the weak invariance model, which in addition to the configural invariance, constrains the loadings of the items on the latent mental wellbeing factor to be the same in the three groups. The third is a strong invariance model which constrains, in addition to the loadings, the intercepts of items to be the same across the three groups. Finally the strict invariance model constrains loadings, intercepts, and the error variances of the items.
Web Table 1 below shows a number of fit measures for the four models. Firstly, the exact fit hypothesis for the configural invariance is rejected. This implies that the most unconstrained model fits data less than perfect to start with. But also note that N is rather large in our case (36,623) so even small misfits may become insignificant. Our aim is not to test or validate the SWEMWBS scale itself, but establish its invariance across the three groups. So we leave a side the misfit of the configural invariance model. We would like to report, however, that adding a covariance between the error terms of item 1 and 2 and between item 5 and 7 improves the model fit rather significantly. Adding these two error covariances reduces the model χ2 to 2268.22 (36) The fit of the weak invariance model is comparable to the configural invariance model. While a likelihood ratio test favors the configural invariance over weak invariance (compare the χ2 values of the two models), this could again be due to very large N. Other fit measures, in fact, indicate that the weak invariance model fits somewhat better than the configural invariance model (e.g. RMSEA, BIC, and TLI). A comparison of item loadings across the three religion groups in the configural invariance model shows that the loadings vary only marginally. The largest differences are for the loadings of item 6 and 7, both of which are somewhat higher among Muslims and Christians than the non-religious. In fact, removing items 6 and 7 from the scale makes the χ2 difference test between the weak invariance and the configural invariance models statistically insignificant. While there is a case to remove these two items from the model, the sizes of the differences in the loadings across the religion groups seem rather minor.
The conclusions are similar if we compare the strong and strict invariance models with the configural invariance model. We thus conclude that the short version of the Warwick-Edinburgh mental well-being scale measures mental well-being rather similarly across the three religion groups with a caveat regarding item 6 and 7.
GHQ Now, we test measurement invariance of the 12-item subjective well-being score (GHQ) across the nonreligious, Christians of any denomination, and Muslims of any denomination. We will use the data from Understanding Society Wave 1. The 12 items ask about (1) concentration (2) loss of sleep (3) playing a useful role (4) capable of making decisions (5) constantly under strain (6) problem overcoming difficulties (7) enjoy day-to-day activities (8) ability to face problems (9) unhappy or depressed (10) losing confidence (11) believe in self-worth (12) general happiness. The same as above, we fit four models with varying degrees of strictness of measurement invariance (see e.g. Kline 2016), namely configual invariance, weak invariance strong invariance and strict invariance models.
Web Table 2 below shows the fit measures of these four models. The exact fit hypothesis for the configural invariance is flatly rejected. This implies that the most unconstrained model fits data poorly to start with. This poor fit could be due to very large N, but even other approximate fit measures (CFI, TLI, and SRMR) indicate rather poor fit. This implies that GHQ may not be measuring a one-dimensional concept. A likelihood ratio test rejects the weak invariance model in favour of the configural invariance model (χ2 (22) = 152.86, P < 0.001). This again could be due to very large N. In fact, other fit measures of the weak invariance model show an improvement of model fit compared to the configural invariance model. RMSEA, BIC, TLI, and CIT for example favour the weak invariance model. This shows that measurement invariance of GHQ across the three religions can be accepted, however, the unconstrained model fits rather poorly.
Web We now modify the configural invariance model to attain a relatively better fitting baseline model. After inspecting the R-squared per item and the differences in loadings across the three religions, we remove items 1, 3, 4, 7, and 8. The R-squared values for these items are respectively 33%, 23%, 27%, 36%, and 34%. These values are too low to justify including them in the same scale. Next, looking at modification indices, we add an error covariance between item 11 and item 12 and between item 2 and item 5. This results in a relatively well fitting modified configural invariance model with 7 items. We now test various invariance models building on this modified configural invariance model.
The fit of the modified weak invariance model is comparable to the modified configural invariance model. While a likelihood ratio test favors the configural invariance over weak invariance (compare the χ2 values of the two models), this could again be due to very large N. Other fit measures indicate that the modified weak invariance model fits better than the configural invariance model (e.g. RMSEA, BIC, CFI and TLI). Also, in the modified configural invariance model the loadings differ only marginally across the religions. Strong and strict invariance models also fit data relatively well. We thus conclude that the reduced version of the GHQ scale measures wellbeing relatively similarly across the three religions.
Web The Figure below shows the effect of religion variables on the modified, reduced versions of GHQ and of SWEMBWS (items 6 and 7 removed). The models rely on complete-case analysis and original GHQ scale (i.e. not-reverse coded as in the main manuscript), and that for simplicity importance and attendance variables were scaled to the [0 1] range and treated as continuous. These effects are similar to the ones with the full versions of GHQ and SWEMBWS. Note that these effects are estimated with a complete-case analysis, hence standard errors tend to be somewhat larger. Figure. Religiosity variables in relation to reduced versions of SWEMWBS and GHQ.        Note: Religiosity measured in Wave 1 (2009/2011) and outcomes in Wave 4 (2012-2014); quantile regression models were used-coefficients are interpreted analogously to linear regression: e.g., Q50 shows the median difference in mental wellbeing comparing those with Muslim compared with no religious affiliation, while Q90 shows the difference at the 90 th quantile (far right-end of the distribution). For simplicity religious attendance and importance were scaled to the [0 1] range and treated as continuous.