|Association between nutritional interventions and depressiona|
|Analysis||Total, n||Cases, n||Nutritional intervention, HR (95% CI)|
|Excluding late cases (>6 years of follow-up)||3,893||211||1 (ref)||0.88 (0.64 to 1.21)||0.73 (0.52 to 1.05)|
|Including only participants with diabetes||1,958||113||1 (ref)||0.71 (0.46 to 1.09)||0.59 (0.36 to 0.98)|
|Including only participants with hypertension||3,181||181||1 (ref)||0.93 (0.66 to 1.31)||0.79 (0.54 to 1.16)|
|Including only participants with hypercholesterolemia||2,697||158||1 (ref)||0.93 (0.64 to 1.35)||0.81 (0.54 to 1.21)|
|Including only participants with obesity||1,781||113||1 (ref)||0.85 (0.55 to 1.31)||0.69 (0.41 to 1.13)|
|Excluding participants with limiting conditionsb||3,074||184||1 (ref)||0.81 (0.57 to 1.13)||0.68 (0.46 to 0.99)|
|Excluding participants with cancer||3,824||216||1 (ref)||0.90 (0.66 to 1.24)||0.79 (0.56 to 1.12)|
|Excluding participants with potentially missing eventsc||3,867||224||1 (ref)||0.91 (0.66 to 1.24)||0.78 (0.55 to 1.10)|
|Multiple imputation for missing values||82,383||4,740||1 (ref)||0.95 (0.69 to 1.32)||0.75 (0.52 to 1.07)|
Abbreviations: HR hazard ratio, MD Mediterranean diet, ref reference group, EVOO extra virgin olive oil.
aAdjusted for age, sex, recruiting center, body mass index, smoking, physical activity during leisure time, educational level, marital status, alcohol and total energy intake, and the presence of several diseases at baseline (cancer, diabetes mellitus type 2, hypertension, hypercholesterolemia, fractures, Parkinson disease and chronic bronchitis).
bSuch as fractures, PD, and chronic bronchitis.
cApproaches for dealing with missing data were as follows. We implemented the suggestions given by Groenwold et al.  to handle potentially missing outcome data in randomized trials. In order to conduct sensitivity analyses on missing outcomes, we considered as missing outcome data the potential occurrence of depression for 59 participants for whom we had no documented incident depression and who were lost to follow-up for 2 years or longer. We conducted two sensitivity analyses: 1) in the complete case analysis we excluded these 59 subjects from the computation of HRs and adjusted the Cox model for all the covariates mentioned in Table 3; 2) in multiple imputations, 20 values were sampled from an estimated uniform distribution (also taking into account the previously mentioned predictors) and imputed for the 59 participants lost to follow-up adding a random term. Hence, 20 datasets with imputed outcomes were created. Each dataset was analyzed using multivariable-adjusted Cox models and, subsequently, the results were pooled by using standard techniques, also taking into account the variation between imputed data sets .
Sánchez-Villegas et al.
Sánchez-Villegas et al. BMC Medicine 2013 11:208 doi:10.1186/1741-7015-11-208