We used listwise deletion to manage missing data (52), but subsequently adopted the Expectation Maximisation (EM) method for purposes of robustness checks (53). Since each approach has its drawbacks (e.g., EM replaces missing values by estimating parameters) (52) adopting both helped validate our results across two disparate methods for managing missing values (54). Rates of missing data ranged from 0% (e.g., age, gender, ethnicity) to 40.6% (antidepressant usage), and 66.1% (HbA1c (mmol/mol)).
Descriptive data is shown in Table 1. Antidepressant users were more likely than non-users to report fair (31.7% versus 14.9%), bad (18.8% versus 3.1%) or very bad (7.9% versus 1.1%) SRH (χ2 (1, N = 6115) = 676.432, p < 0.001), and experience poorer cardiometabolic health in relation to waist/hip ratio (cm) (M = 0.893 versus 0.877) (t(5036) = 3.728, p < 0.001), BMI (> 30kg/m2) (M = 30.156 versus 25.883) (t(5317) = 14.452, p < 0.001), diastolic blood pressure (mmHg) (M = 72.950 versus 70.452) (t(5463) = 5.224, p < 0.001), systolic blood pressure (mmHg) (M = 125.249 versus 122.036) (t(5463) = 4.309, p < 0.001) and HbA1c (mmol/mol) (M = 41.851 versus 39.335) (t(487.019) = 4.781, p < 0.001).
Antidepressant users were less likely than non-users to be HDL-C deficient (mmol/L) < 0.9 (38.6% versus 48.4%), (χ2 (1, N = 6115) = 21.302, p < 0.001), and have elevated waist/hip ratio (> 0.85cm) (61.4% versus 48%), (χ2 (1, N = 6115) = 40.296, p < 0.001), BMI (> 30kg/m2) (37.1% versus 20%), (χ2 (1, N = 6115) = 96.088, p < 0.001), HbA1c (> 48 mmol/mol) (9.3% versus 3.5%), (χ2 (1, N = 8304) = 46.979, p < 0.001), diastolic blood pressure (>90 mmHg) (6.4% versus 3.9%) (χ2 (1, N = 6115) = 8.696, p = 0.003), and systolic blood pressure (>140 mmHg) (19.5% versus 13.2%) (χ2 (1, N = 6115) = 18.588, p < 0.001), and be diagnosed with hypertension (34.4% versus 19.2%) (χ2 (1, N = 6112) = 77.719, p < 0.001).
……………………………………………..
[Insert Table 2 about here]
Final regression models predicting metabolic factors from antidepressant usage, self-rated health, and metabolic covariates.
………………………………………………
- Bootstrapped hierarchical multiple regression analysis
Bootstrapped hierarchical logistic regression analysis was used to determine if accounting for SRH negates a previously significant association between antidepressant usage and a metabolic biomarker, suggesting a possible mediator effect (44). Structural equation modelling (SEM) was later used to explore suspected mediator pathways. In each regression analysis we predicted a specific metabolic outcome (e.g., HDL-C), controlling for all other metabolic variables. We assembled four models in each regression analysis: Model 1 (metabolic outcome = Intercept + Age + Gender + Socio-economic Class + Ethnicity), Model 2 (metabolic outcome = Intercept + Age + Gender + Socio-economic Class + Ethnicity + Antidepressant usage), Model 3 (metabolic outcome = Intercept + Age + Gender + Socio-economic Class + Ethnicity + Antidepressant usage + SRH), and Model 4 (metabolic outcome = Intercept + Age + Gender + Socio-economic Class + Ethnicity + Antidepressant usage + SRH + metabolic covariates). Thus, SRH was added to the equation (Model 3) after first accounting for antidepressant usage (Model 2).
Table 2 shows hierarchical bootstrapped regression estimates and overall model parameters. Antidepressant usage significantly predicted HDL-C (mmol/L) (Model 2) (β = -0.053, p = 0.002), increasing the explained variance, ∆R2 = 0.003, F (1, 2967) = 9.684, p = 0.002. Users reported lower HDL-C (mmol/L) levels compared to non-users. However, adjusting for SRH (Model 3) completely negated the association between antidepressant usage and HDL-C (mmol/L) (β = -0.011, p = 0.514). SRH significantly predicted HDL-C (mmol/L) (β = -0.150, p = 0.000), accounting for an additional 1.9% of the variance in HDL-C (∆R2 = 0.019, F (1, 2966) = 70.121, p = 0.000). Poorer SRH was associated with lower HDL-C (mmol/L) levels. Addition of other metabolic covariates (Model 4) markedly increased the explained variance (∆R2 = 0.105, F (5, 2961) = 87.019, p = 0.000), but did not negate the significant relationship between SRH and HDL-C (mmol/L) (β = -0.150, p = 0.000).
Use of antidepressants was significantly associated with waist/hip ratio (cm) (Model 2) (β = -0.085, p = 0.000), contributing marginally to the explained variance (∆R2 = 0.007, F (1, 2967) = 37.240, p = 0.000). Users reported higher central adiposity compared to non-users. Adjusting for SRH (Model 3) failed to nullify the relationship between antidepressant usage and waist/hip ratio (cm) (β = 0.045, p = 0.002). SRH was associated with waist/hip ratio (cm) (β = 0.145, p = 0.000), contributing an additional 1.8% of the variance in central adiposity (∆R2 = 0.018, F (1, 2966) = 99.302, p = 0.000); poorer SRH was related to higher waist/hip ratio (cm) scores. Including other metabolic covariates (Model 4) explained almost 15% an additional variance (∆R2 = 0.146, F (5, 2961) = 222.096, p = 0.000), negating the association between antidepressant usage and waist/hip ratio (cm) (β = 0.019, p = 0.127). However, SRH remained a significant predictor (β = 0.050, p = 0.000).
Antidepressant usage significantly predicted BMI kg/m2 (Model 2) (β = 0.120, p = 0.000), improving the percentage variance explained (∆R2 = 0.014, F (1, 2967) = 42.485, p = 0.000). Users had higher BMI kg/m2 scores, compared to non-users. Accounting for SRH (Model 3) did not markedly affect the association between antidepressant usage and BMI kg/m2 (β = 0.063, p = 0.001). SRH significantly predicted BMI kg/m2 (β = 0.205, p = 0.000), explaining an additional 3.6% of the variance (∆R2 = 0.036, F (1, 2966) = 114.406, p = 0.000). Poorer SRH was associated with higher BMI kg/m2. Including other metabolic covariates (Model 4) explained significant additional variance (∆R2 = 0.272, F (5, 2961) = 241.678, p = 0.000), attenuating the contribution of antidepressant usage (β = 0.033, p = 0.038), but not SRH (β = 0.089, p = 0.000).
There was no link between use of antidepressants and diastolic blood pressure (mmHg) (Model 2) (β = 0.032, p = 0.083), with no improvement in the explained variance (∆R2 = 0.001, F (1, 2967) = 3.004, p = 0.083). SRH (Model 3) also failed to predict diastolic blood pressure (mmHg) (β = 0.048, p = 0.015) (∆R2 = 0.002, F (1, 2966) = 5.877, p = 0.015). Inclusion of the various metabolic covariates (Model 4) significantly improved the explained variance, by an additional 43% (∆R2 = 0.430, F (5, 2961) = 454.723, p = 0.000). Antidepressant usage also failed to predict systolic blood pressure (mmHg) (Model 2) (β = 0.016, p = 0.328) (∆R2 = 0.000, F (1, 2967) = 0.955, p = 0.328). Accounting for SRH (Model 3) did not improve the model (β = 0.028, p = 0.115). (∆R2 = 0.001, F (1, 2966) = 2.486, p = 0.115). Addition of metabolic covariates (Model 4) explained an additional 33.4% of the variance in systolic blood pressure (mmHg) (∆R2 = 0.334, F (5, 2961) = 408.529, p = 0.000).
Use of antidepressants significantly predicted HbA1c (mmol/mol) (Model 2) (β = 0.088, p = 0.000), accounting for a marginal increment in explained variance (∆R2 = 0.007, F (1, 2967) = 24.814, p = 0.000), over and beyond sociodemographic factors. Antidepressant users had higher HbA1c (mmol/mol) levels than non-users. Adjusting for SRH (Model 3) attenuated the previously significant association between antidepressant usage and HbA1c (mmol/mol), whereby the β value showed a 44.318% decrease (β = 0.049, p = 0.007), and was no longer significant at the bonferroni adjusted alpha level of p < 0.005. Poorer SRH predicted higher HbA1c (mmol/mol) levels (β = 0.138, p = 0.000), accounting for an additional 1.6% of the variance (∆R2 = 0.016, F (1, 2966) = 55.487, p = 0.000. Adding other metabolic covariates (Model 4) marginally improved the explained variance (∆R2 = 0.043, F (5, 2961) = 30.897, p = 0.000), further attenuating the association between antidepressant usage and HbA1c (mmol/mol) (β = 0.038, p = 0.033). SRH remained significant (β = 0.088, p = 0.000).
- Structural equation modelling
The attenuated association between use of antidepressants and HDL-C (mmol/L) after accounting for SRH suggested a possible mediator effect for SRH (44). Thus, SEM (IBM AMOS SPSS, version 28) was used to explore this indirect association (45). Model fit was assessed using the following criteria: model chi-square χ2 (CMIN) (p > 0.05), χ2 (CMIN)/df < 2.00, root mean square error of approximation (RMSEA) < 0.07, comparative fit index (CFI) ≥ 0.95, Tucker and Lewis Index (TLI) ≥ 0.95 and normed fit index (NFI) ≥ 0.95 (55). We tested an initial model in which antidepressant usage was allowed to affect HDL-C (mmol/L), both directly and indirectly, mediated by SRH; to test for bidirectional causality, HDL-C (mmol/L) was also allowed to affect antidepressant use, directly and indirectly, mediated by SRH. We used the specification-search function, which evaluates the fit of different candidate models. This generated one good-fitting model based on the BIC (Bayesian Information Criterion) value (0), and the χ2 (CMIN)/df value. This model is presented in Figure 2. Multiple fit criteria for this model indicated good fit: χ2 (CMIN) = 0.294, df = 1, p = 0.587, χ2 (CMIN)/df = 0.294, RMSEA = 0.000, CFI = 1.000, TLI = 1.006 and NFI = 1.000. The model revealed an intermediary role for SRH. Both antidepressant usage and HDL-C (mmol/L) predicted SRH; poorer SRH was associated lower HDL-C (mmol/L) levels (β = -0.302, p = 0.001) and antidepressant use (β = 0.976, p = 0.001). There was no direct association between HDL-C (mmol/L) and antidepressant usage. Squared multiple correlations showed the default model explained 11.7% of the variance in SRH.
……………………………………………..
[Insert Table 3 about here]
Final regression models predicting metabolic factors from antidepressant usage (past 7 days), self-rated health, and metabolic covariates.
………………………………………………
.……………………………………………..
[Insert Table 4 about here]
Final regression models predicting metabolic factors from antidepressant usage, self-rated health, and metabolic covariates, using log-transformed data.
………………………………………………
.……………………………………………..
[Insert Table 5 about here]
Final regression models predicting metabolic factors from antidepressant usage, self-rated health, and metabolic covariates, using estimated values based on the EM (Expectation-Maximisation) algorithm.
………………………………………………
In addition to using bootstrapped regression analysis, which helped validate the statistical findings by repeating the analysis across 1000 random samples, taken from original sample (56), we also performed eight additional robustness checks to see how the results are affected by various sources of uncertainty, including variations in methods of analysis, definitions of variables, analytic protocol, and missing data (57).
First, we repeated the analysis with an alternative measure of antidepressant usage, as an additional validation. Specifically, we examined use of antidepressants (at least one drug) specifically during the past 7 days. The emerging bootstrapped regression estimates generally supported the initial findings, whereby adjusting for SRH (Model 3) attenuated previously significant associations of antidepressant usage (in the past 7 days) with both HDL-C (mmol/L) (β = -0.008, p = 0.659), and HbA1c (mmol/mol) (β = 0.045, p = 0.012), highlighting the robustness of the initial results (see Table 3). One exception was a now attenuated association between antidepressant usage (in the past 7 days) and waist/hip ratio (cm) (β = 0.035, p = 0.031), after controlling for SRH.
Second, we repeated the above main analysis using SRH collapsed into a dummy variable. There has been some debate regarding whether SRH categories represent an arbitrary classification of underlying continuous phenomena, or intrinsically distinct psychological states (58). Assuming the latter, we felt it was important to determine if the present findings are sensitive to SRH categorisation. This approach generated partially identical results whereby adjusting for SRH (Model 3), as expected, attenuated the association of antidepressant usage with HDL-C (mmol/L) (β = -0.021, p = 0.219), but not HbA1c (mmol/mol) (β = 0.057, p = 0.002).
Third, we repeated the above analyses using a different measure of HbA1c (percentages rather than mmol/mol). This produced identical results, again with adjustment for SRH (Model 3) only attenuating the influence of antidepressant usage on HDL-C (mmol/L) (β = -0.021, p = 0.219).
Fourth, we repeated the analyses controlling for use of any prescribed mental health medications in the past 7 days (Model 1). Unsurprisingly, this negated the contribution of antidepressant usage across all outcome variables (Model 2) (all p’s > 0.005) since antidepressants are likely to be included in the medication being used.
Fifth, we conducted curvilinear regression analyses (59), to see how quadratic functions of the metabolic covariates (added as Model 5) affect observed contributions of antidepressant usage (Model 2) and SRH (Model 3). Results for SRH and antidepressant usage were largely unchanged, despite quadratic functions significantly improving the explained variance, across all regression models (Model 5) (all p’s < 0.001). SRH remained a significant predictor of HDL-C (mmol/L) (β = 0.077, p < 0.001), waist/hip ratio (cm) (β = -0.047, p < 0.001), BMI kg/m2 (β = -0.078, p < 0.001), and HbA1c (mmol/mol) (β = -0.078, p < 0.001), weakening the relationship between antidepressant usage and HDL-C (mmol/L) (β = -0.021, p = 0.219).
Sixth, we conducted the regression analysis using log-transformed data (common logarithm – to base 10) for metabolic factors, partly due to the positive skew of some variables (60). This confirmed the above patterns, with SRH (Model 3) primarily negating the association between antidepressant usage and HDL-C (mmol/L) (β = -0.008, p = 0.658) (see Table 4).
Seventh, we re-ran the original regression analyses using a different approach for dealing with missing data (EM rather than list-wise deletion). This analysis generated roughly identical results, with addition of SRH (Model 3) attenuating the covariance between antidepressant usage on HDL-C (mmol/L) (β = -0.015, p = 0.302) (see Table 5). One notable anomaly in the output was a previously non-significant association between antidepressant usage and diastolic blood pressure (mmHg) (Model 2), which was now significant (β = 0.044, p = 0.003). However, as in the original analysis, neither antidepressant usage nor SRH emerged significant in the subsequent block of predictors (Model 3).
Finally, we ran SEM analyses using just the log-transformed data, and then again using both the log-transformed data and alternative measure of antidepressant use (i.e., usage during the past 7 days). This generated models with good fit. Consistent with the original SEM analyses, SRH served as an intermediary, being predicted by both antidepressant usage and HDL-C (mmol/L). The first model accounted for 12% of the variance in SRH (χ2 (CMIN) = 0.313, df = 1, p = 0.576, χ2 (CMIN)/df = 0.313, RMSEA = 0.000, CFI = 1.000, TLI = 1.006 and NFI = 1.000) (see Figure 3a), while the second explained 10.9% (χ2 (CMIN) = 0.031, df = 1, p = 0.860, χ2 (CMIN)/df = 0.031, RMSEA = 0.000, CFI = 1.000, TLI = 1.011 and NFI = 1.000) (see Figure 3b).
We then repeated the SEM analyses using the EM method for managing missing data. This produced two good-fitting models, again with SRH playing an intermediary role, in both models. However, whereas in the original analyses, SRH was the outcome variable, predicted by antidepressant use and HDL-C (mmol/L), the first EM model showed SRH predicting both variables (see Figure 3c), while in the second model HDL-C (mmol/L) predicted SRH, which in turn predicted antidepressant usage (see Figure 3d). Both models generated roughly identical estimates (χ2 (CMIN) = 1.071, df = 1, p = 0.301, χ2 (CMIN)/df = 1.071, RMSEA = 0.003, CFI = 1.000, TLI = 1.000 and NFI = 0.999).
……………………………………………..
[Figure 3a - Structural equation model using log-transformed data]
[Figure 3b - Structural equation model using antidepressant usage (past 7 days), and log-transformed data]
[Figure 3c - Structural equation model using estimates from the expectation-maximisation algorithm (first model)]
[Figure 3d - Structural equation model using estimates from the expectation-maximisation algorithm (second model)]
[Legend: ADU (antidepressant usage), SRH (self-rated health), and HDL-C (high density lipoprotein, “good” cholesterol). All SEM models were derived using the specification search function available on IBM SPSS Amos]
………………………………………………