There are two primary endpoints in this study based on comparing conditions on:

1.       Change in pts' LDL-cholesterol after 12 months in the study.

2.       Difference in the level of SFA (% energy) after 12 months in the study.

Additional analyses will be conducted for secondary hypotheses relating to: 1. Difference in total fat intake (% energy) after 12 months in the study; 2. Change in pts' mean body weight; and 3. Difference by condition in the number of intervention steps implemented by physicians (as measured by PEIs).

The analyses will be developed by Dr. Stanek in close collaboration with Drs. Hebert and Ockene, and implemented both at the Umass-Amherst campus and at UMMC.  Interchange of analysis results, data sets, analysis programs, and reports will be facilitated by frequent face-to-face meetings, and electronic mail.

As discussed in D.9., measures will be put into place to optimize the rate of data completion. Nonetheless there will be some pts with incomplete data. Descriptive analyses will characterize study completers, and be augmented with ANOVA and chi-square tests as appropriate. Logistic regression analyses will be conducted to evaluate differences between completers and non-completers, focusing on factors that are differentially distributed between conditions. Particular attention will be given to gender and education differences, since these factors appeared to be potential effect modifiers in the WATCH study. Because of missing data, different subsets of pts may be eligible for analyses of LDL-C and SFA (%energy).  Complete analyses will be conducted separately for these endpoints. Variables found to both differ between completers and non-completers and to adhere to the assumptions of the regression models will be included as covariables in subsequent primary and secondary analyses to control for possible selection bias. Analyses also will be conducted that allow for unbalanced and missing data, with the logistic regression analyses serving to help guide interpretation of these results.

Change in LDL-C levels (baseline minus one year) will be constructed for each pt and serve as the primary outcome variable for testing hypothesis #1. Since the study design is stratified by physicians, we will conduct primary analyses using physicians as a fixed blocking factor in the analysis.  All analyses will make use of PROC MIXED in SAS due to the expanded flexibility of the mixed modeling paradigm.  The principal comparison will control for physicians and other important selection variables identified at baseline, while testing for differences in condition. An important covariable that will be included is use of cholesterol-lowering medications (started after baseline). The rate of such medication used, as modified by the intervention, as well as the effect on change scores will be developed, with analyses possibly stratified on medication use.  Models will be compared using likelihood ratio tests based on maximum likelihood estimates, with final model parameters estimated via REML. Subsequent analyses will consider possible effect modification by education, gender, and their interaction.

In addition to the primary analyses using physicians as fixed blocking factors, additional analyses will consider physicians as a random factor, enabling extrapolation of intervention effects to a broader group of physician practices.  These analyses will make use of broad inference as described by McLean et al., and summarized by Little et al. (130, 131) Attention will be paid to interpreting the differences between results based on fixed and random effects, so as to distinguish results specific to this randomized trial from results generalizable to other settings.

Additional analyses will be conducted that include the baseline LDL-C value.  Because patient eligibility is established via duplicate fingerstick cholesterol measures and baseline LDL-C levels are taken from duplicate venous measures, there should be minimal concern over regression to the mean. Consequently, including the baseline LDL-C value should result in a more powerful test. Although little regression to the mean is anticipated, some regression is possible since due to practical constraints, the second finger stick measure and the first venous LDL-C measures are conducted on the same day (i.e., within 1 hour of each other).  Analyses will be developed that identify this possible regression effect.  While not affecting the primary comparison between conditions, if regression to the mean is present, it will affect condition-specific estimates.  Alternative analysis approaches will be developed if regression to the mean is detected; e.g., analyses that account for the regression, or ones that use only the second venous LDL-C baseline measures.

Models for SFA (% energy) at one-year will be constructed in a manner similar to analyses for LDL-C.  Large day-to-day variability within pts results in more powerful analyses that compare group differences between SFA (% energy) at one year, as opposed to change scores when based on 24HR recall data.  Randomization to intervention groups guarantees comparability of pts in conditions, with imbalances due to non-response accounted for by inclusion of additional covariables identified via the initial logistic regression analyses. However, the power of these analyses may be augmented by inclusion of pt specific covariates.  Such covariates will include education, gender, age (both continuous and categorical) and other patient factors. Model development will proceed as for LDL-C, with physicians treated first as fixed block effects, and then considered as random factors for generalization of study results to other settings.

Models for secondary hypotheses will be developed in a similar manner to analyses for primary hypotheses.  Models for total fat intake will parallel analyses for SFA (% energy).  Analyses of change in body weight will make use of (baseline - one year) change scores, similar to analyses for LDL-C.  Analyses of the number of intervention steps implemented by physicians will be conducted using the PEI scores as the outcome variable. When considering PEI scores as an outcome, a covariable will be constructed corresponding to the time-weighted number of patients previously receiving the condition II intervention to investigate possible transfer of the systems-based approach to altering behavior among the condition I intervention pts. Additionally, PEI score will be fit as a potential effect modifier in analyses of the primary outcome variables, LDL-C and SFA. 

Additional analyses will be conducted to address other ancillary research questions.  Change in fat intake based on the 7DDR will be fit to the Keys and Hegsted equations (65) as we have done previously (79) to predict change in total cholesterol, thus enabling us to evaluate how well, on average, the study patients report their dietary change.  Correlation analyses will be conducted to examine whether changes in intake of SFA are independent of intake of other fatty acids; and whether reduction in SFA will predict reductions in caloric density of the diet as well as in total caloric intake.  Analyses will be conducted to examine the effect of the GNI on lipid and dietary outcomes. Predicted values of one-year LDL-C levels will be developed from the final fixed effect LDL-C models and used to evaluate the percent of pts achieving their LDL-C goals, and compare these percents between conditions.