Our main analysis model (Main) therefore included age, sex, education, BMI, LDL-C, HDL-C, smoking status, pack-years of smoking, ETS, physical activity, indicator variables for summer and fall (winter and spring served as reference), humidity, plus city, which was included to capture spatial (unmeasured) confounding.
We first describe the modelling framework for our main analysis, model A. This model brings together IPD and AD into the same synthesis model, with both types of evidence contributing to the estimation of all key models parameters (i.e. absolute and relative effectiveness estimates).
To evaluate the robustness of our main analysis model, we performed a series of models that included additional covariates.
The main analysis modelling was limited to 10 years, in the absence of robust longer-term risk models.
The main analysis model was modified in such a way that the individual EQ-5D dimension was employed instead of the EQ-5D Index.
For our analysis, the imputation model will include all variables which we believe may contain information about the missingness mechanism at 12 and 15 months and must include all variables that will be used in the main analysis model [ 66].
Finally, due to the relatively high numbers of missing on the GMH index, we reran the main analysis (Model 4, in Analysis 1, 2 and 3) with the original, unimputed GMH index to check the stability of the results.
The remaining variables were entered into the main analysis model in six steps: demographic factors; life time cardiovascular and affective conditions; the additional explanatory value of other health related indicators; contextual factors; two-way interaction terms; three-way interaction terms.
Furthermore we added covariates known to be associated with cardiovascular disease or with systemic inflammation such as hypertension, diabetes, and intake of statins to investigate the robustness of our main analysis model.
