2.10. Statistical Analysis

Descriptive analysis to compare the baseline characteristics (including intake assessed by baseline FFQ) of those in the predictive validation group using FFQ and LDL-C data to the concurrent validation subset that completed the FFQ and 7DDR was conducted by using Student t-test for continuous variables and chi-square test for categorical variables.

For the predictive validity analysis, we used multiple linear regression and Pearson correlation coefficients to examine the association and correlation of change in the PDS with concomitant change in LDL-C, and logistic regression to examine the odds of having a 5% reduction in LDL-C with change in the PDS. In all models, we used robust variance estimates to account for intra-cluster correlations (n = 94 two-member households; n = 3 three-member households) to account for households randomized together in the main trial. We adjusted for the following covariates in the fully adjusted model: sex (male/female), age (continuous), baseline LDL-C (continuous), energy intake (continuous), use of cholesterol-lowering medications (yes/no), ethnicity (Caucasian/Asian/African and Caribbean/Other/Unknown), BMI (continuous), smoking status (current/never and past), family history of CVD (yes/no), education (high school or less/undergraduate and college/graduate), physical activity level (low/medium/high) and intervention group (control/intervention). These same methods were repeated in our sensitivity analyses to assess the effect of median and quintile cut-offs, and weighting of the saturated fat/cholesterol component during the PDS development.

For the concurrent validation analyses, we calculated Pearson correlation coefficients between the PDS from the FFQ and the PDS from the 7DDR, with and without energy adjustment (using the residual method) as absolute score (6–30) and as quantiles (1–4). We further calculated deattenuated correlation coefficients between the PDS computed from the FFQs and 7DDR, correcting for within-person error in the 7DDR, using the following formula:

where P_c is the corrected correlation coefficient between the FFQ and 7DDR, Po is the observed energy-adjusted correlation coefficient between the FFQ and 7DDR, γ is the ratio of within-to-between-person variation in the 7DDR PDS, and k is the number of diet record days recorded (k = 7 in this instance).

We also calculated deattenuated Pearson or Spearman correlation coefficients between the single components (as quantiles and absolute amounts) of the PDS from the FFQ and from the 7DDR, as well as energy (kcal) and macronutrients (carbohydrate, protein, fat). The single components (absolute amounts) of the PDS were log transformed after adding (0.0001 servings/day) to the data to allow the zeros to be fixed to a non-zero value as the components were positively skewed. We additionally analysed the absolute agreement between the two methods (FFQ and 7DDR) with the Bland–Altman method, which determines the average agreement between two methods by calculating the mean of their differences (FFQ-7DDR) against the mean intake of the two measures (FFQ + 7DDR/2). The 95% limits of agreement (LOA) provide an interval within which 95% of these differences are expected to fall.

Lastly, we calculated the proportion of participants correctly categorized (same quantile) and grossly misclassified (opposite quantile) for the PDS and single components. Statistical tests were 2-sided and p < 0.05 was considered statistically significant. The statistical analyses were conducted with Stata statistical software (Stata Statistical Software: Release 15. College Station, TX, USA).