Study population

Details on the PREDIMED design and methods are described in detail elsewhere [18, 23, 43]. Briefly, the PREDIMED trial was a multicenter, randomized, controlled, single-blinded cardiovascular disease prevention trial conducted in Spain [41]. It was designed to assess the effects of the MedDiet on cardiovascular disease in 7,447 participants recruited between 2003 and 2009. Eligible participants were aged 55 to 80 and at high risk for developing cardiovascular disease (CVD). High risk was defined as having type 2 diabetes or at least three of the following major (CVD) risk factors: current smoking, hypertension, elevated low-density lipoprotein cholesterol levels, BMI ≥25 kg/m², or a family history of premature coronary heart disease (CHD). After providing written informed consent, participants were randomized to one of three interventions; a traditional MedDiet supplemented with either complementary extra-virgin olive oil (EVOO) or tree nuts or a control diet (advice to reduce all types of dietary fat). The control group was excluded from the present analyses because the focus of this study was the adherence with the intervention promoting the MedDiet. Inclusion criteria into this study are depicted in the flow diagram figure provided below, along with the CONSORT checklist for randomized trials. The trial ended in 2010, after a median follow-up of 4.8 years, because of the observed benefit of the MedDiet compared to the low-fat control diet in the prevention of CVD. Institutional Review Boards at 11 recruiting centers approved the study protocol [44]. No harm or unintended effects were reported in any arm [41].

When one-year dietary adherence was the outcome of interest, the present study excluded participants missing information on any of the 14-point dietary adherence score items at one year of follow-up (n = 799), leaving 4,198 participants for analysis. When four-year dietary adherence was the outcome of interest, participants recruited after November 2006 (n = 1,495) were excluded because subsequent follow-up was less than four years. Participants who had missing information on any of the 14-point dietary adherence score items at four years of follow-up were further excluded (n = 1,149), leaving 2,353 participants available for analysis.

Outcome assessment

Registered dietitians conducted quarterly group sessions and one-on-one in-person interviews to deliver a comprehensive motivational educational intervention aimed at modifying participant eating habits. Dietitians collected detailed dietary information at baseline and yearly thereafter during follow-up. Individual interviews and group sessions were conducted every three months throughout the trial. A previously validated 14-item Mediterranean Diet Assessment Tool [45] was the primary method for assessing adherence to the intervention (Additional file 1: Figure S1). PREDIMED dietitians assessed participant adherence using this tool during each visit. A value of 0 (non-compliant) or 1 (compliant) was assigned to each item [46]. Higher scores reflected better adherence. High adherence was defined as meeting at least 11 of the 14 items. This cut-point was used because roughly half of participants complied with 11 or more items at each follow-up visit.

Covariate assessment

Dietitians administered a validated 137-item food frequency questionnaire (FFQ) at each yearly visit [47], from which total energy and alcohol intake was computed [48]. Another questionnaire collected information on sociodemographic variables, lifestyle, and family history data. A validated Spanish version of the Minnesota questionnaire [49, 50] was used to assess physical activity. Investigators reviewed medical records at baseline and yearly thereafter to assess medical diagnoses. Nurses measured weight and height using standardized procedures, and blood pressure using a validated [51] semiautomatic oscillometer in triplicate (Omron HEM_705CP). Primary care doctors assessed participants for new diagnoses of hypercholesterolemia, hypertension and type 2 diabetes. Definitions for these diagnoses are described elsewhere [42].

Exposure assessment

Potential baseline predictors of adherence were assessed based on clinical relevance and findings from previously published studies. These included sex (male, female), age (<65 years, ≥65 years), highest educational level attained (less than primary school, primary school, secondary school, university or more), occupation (retired, working, housewife, unemployed or unable to work), marital status (married, other), number of people in household, continuous cardiovascular risk factor score (one point was assigned for each of the following diagnoses or conditions: type 2 diabetes, hypertension, high blood cholesterol, family history of premature CHD, depression, and obesity; scores of 0–1 and 5–6 were collapsed due to low frequency), individual cardiovascular risk factors (each of the six factors included in the cardiovascular risk score were also assessed separately instead of as a continuous score), systolic blood pressure and diastolic blood pressure (SBP and DBP, continuous, per 5 mmHg), waist circumference (continuous, per 5 cm), physical activity (tertiles of MET-min/day), smoking status (never, former, current), total energy intake (quartiles of kcal/day), alcohol other than wine (low: <10 g/day for men, <5 for women; moderate: 10–50 for men, 5–10 for women; high: >50 for men, >10 for women), baseline 14-point dietary adherence score (<11 points, ≥11 points), and MedDiet intervention group (mixed nuts, EVOO). This analysis also evaluated whether the “total workload”, measured in person-years of follow-up and an indicator of how many participants a given center delivers the intervention to throughout follow-up, was associated with dietary adherence.

Statistical methods

Chi-square tests were used to assess differences in distributions of baseline characteristics between those with low adherence (<11 points) and high adherence (≥11 points). In Table 2, crude and multivariate-adjusted logistic regression was used to calculate odds ratios (ORs) of adherence to the MedDiet at one and four years of follow-up according to baseline characteristics. Multivariate models were adjusted for all potential predictors of dietary adherence listed above. To calculate p-values for trend, the median value was assigned to each category and the resulting variable was treated as continuous. Quantiles were not treated as ordinal variables to account for the fact that the differences in median values across quantiles were not equal.

Several sensitivity analyses were conducted using multivariate logistic regression. Analyses looked at associations between potential baseline predictors and adherence at two and three years of follow-up (instead of one and four years), and with ≥10 and ≥12 as alternative cut-points for dietary adherence (instead of ≥11).

All p-values are two-tailed. Values of ≤0.05 are considered statistically significant. All statistical analyses were performed using Stata software (version 12.0, StataCorp 2011, College Station, TX, USA).

Short term dietary adherence (one year of follow-up)

Long term dietary adherence (after four years of follow-up)

Medium-term adherence (two and three years of follow-up)

Additional file 2: Table S1 shows results for the association between potential predictors and adherence at the alternate time points of two and three years of follow-up. All characteristics that predicted lower adherence at both one and four years in the primary analysis multivariate logistic regression models (type 2 diabetes diagnosis, lower physical activity, lower total energy intake, lower 14-point adherence score, randomization to the MedDiet + EVOO arm, and belonging to a PREDIMED center with a lower workload) also predicted low adherence at both two and three years.

Additional file 2: Table S3 presents results from logistic regression analyses of the association between MedDiet intervention (nuts or EVOO) and dietary adherence to nut and olive oil items on the 14-point dietary adherence score (≥4 tbsp olive oil per day; olive oil as main culinary fat; ≥3 servings of nuts per week). Those in the MedDiet + EVOO intervention arm had significantly higher odds of complying with either of the two olive oil items (5 to 10 times the odds) at both one and four years of follow-up. In contrast, those in the nut intervention group had about 20 times the odds of complying with the nut item.

Study design

It is not surprising that the total workload (measured in person years) at a PREDIMED center was associated with both short-term and long-term adherence; the workload likely represents the level of experience the research team had with intervention delivery. Similar findings have been observed in hospitals, where quality of care is often related to number of administered procedures [52]. This finding suggests that multicenter interventions should recruit participants to fewer centers with more participants in each, instead of more centers with fewer participants in each, to maximize effectiveness and adherence. Streamlining intervention delivery would have an added benefit of reducing costs. While this would not explain the difference in adherence, this would free up resources for increased support for participants at risk of poor or suboptimal adherence.

Participants randomized to the MedDiet + EVOO (compared to tree nuts) had lower dietary adherence. This is probably because olive oil is a staple ingredient in the Spanish diet; participants consume olive oil regardless of supplementation from PREDIMED. Nut consumption is not as commonplace. As a result, it is easier for the nut group to adhere to the olive oil criteria compared to the olive oil group’s ability to adhere to the nut criteria. Additional file 2: Table S3 shows that intervention group is a much stronger predictor of complying with the nut adherence item compared to the olive oil adherence items. This suggests that, for dietary interventions providing participants with complementary food items, it may be most effective to provide them with foods that are less commonplace.

Baseline health and lifestyle characteristics

In the present study, many predictors of low adherence with the MedDiet are indicators of poorer baseline health, including various cardiovascular risk factors, less physical activity, and poorer baseline diet. These results are consistent with previous findings investigating predictors of adherence with dietary interventions for reducing fat [29] and carbohydrates [28], family-level interventions [34], and MedDiet interventions [35, 37]. Baseline health status may indicate how much a person values his or her health, which may moderate one’s motivation to comply with the intervention. Alternatively, some research suggests that individuals may be more willing or motivated to make dietary and lifestyle improvements following a medical diagnosis [53]. These findings do not necessarily contradict this notion, as many of these indicators of baseline health are likely long-standing conditions and/or habits; the time during which one is more motivated to make improvements may have passed. Regardless, unhealthy individuals have a greater need for dietary improvement. Thus, they are often the most important targets of dietary interventions. Personalized, higher-intensity intervention approaches may be needed to achieve optimal adherence among less healthy individuals.

Demographic characteristics

Like this study, most previous studies [35, 37] found that women have lower adherence than men. The only exception was a family intervention study [34]. It is possible that because mothers traditionally plan family meals, they were motivated to set a positive example through intervention adherence [34]. However, in this study population, it is possible that spouses and children influence meal preparation, leading to these disparate findings. Different strategies likely have different levels of effectiveness based on sex [54]. However, in the present study the female sex only predicted lower adherence at one year of follow-up, and not four years.

There have been conflicting findings about the relationship between age and dietary adherence [27, 29, 37]. The age range in the present study was restricted to 55–80 years; hence little age variability likely limited the ability to detect an association.

There was little evidence for an association between educational attainment and intervention adherence. While participants with less than a primary school education had lower long-term adherence than those with university level or higher, this did not hold in several sensitivity analyses. Previous studies have found that higher socioeconomic status predicted better dietary adherence, but findings did not hold for long-term adherence [27, 29] and were limited to low-fat dietary interventions [33]. This suggests that dietary interventions may be able to overcome the socioeconomic disparities that often exist in nutrition [55].

Further discussion

The present study has several strengths. First, the sample size was large and it was conducted in an established, long-term, and successful randomized trial. Second, because all study participants were at high risk for CVD, it was possible to assess adherence among participants who were less healthy compared to the general population. Because they also were likely to have poorer baseline diets, adherence was probably especially challenging for these individuals. Thus, significant predictors of adherence may be even more meaningful in this setting. Third, this is one of the few studies that has been able to assess long-term dietary adherence. This is critical, as long-term, high-quality dietary pattern is the relevant dietary exposure for the prevention of chronic disease. Fourth, mutually adjusting for a wide array of baseline characteristics minimized residual confounding. Lastly, significant measurement error is unlikely because only 0.3 % of covariate values were missing, a validated measure for assessing dietary adherence was used [46], and previous analyses show that self-reported dietary intake is highly correlated with biomarkers in this population [18, 23].

It is important to note that because the high adherence is not rare, the ORs do not approximate risk ratios (RRs) and thus should not be incorrectly interpreted as RRs. However, provided appropriate interpretation, ORs still provide valid estimates, and it is more appropriate to apply OR estimates to all individuals within a population. Furthermore, because an OR incorporates both success and failure symmetrically, it is less arbitrary than a RR and thus a more robust estimate [56].

There are also limitations in this study. The potential for measurement error always exists. To include as many people as possible in the present analyses, missing covariate values were imputed for 0.3 % of values. Recall bias, social desirability bias and differential misreporting are always possible when diet is self-reported. Finally, it is always possible that failure to control for unmeasured confounders may have distorted results for predictors of dietary adherence. However, analyses were adjusted for a wide array of important baseline characteristics, and a strong confounder unrelated to these characteristics is unlikely. This unique population of older Spanish participants at high risk for cardiovascular disease may have low generalizability to the general public at lower risk of CVD.

The relative success of a dietary intervention to induce changes in the overall food pattern has been more frequently ascribed to strategies related to negotiation, goal setting, self monitoring, and skill building [57–59]. Other strategies such as the training of dietitians, length and intensity of intervention, frequency of contacts, multiplicity of channels used for the delivery of the intervention, the initial motivation of participants for adherence, and the provision of appropriate means for feedback should not be forgotten.

It is clear that certain participants have greater difficulty complying with dietary interventions. Our results identify specific baseline characteristics that predict better adherence, which is an instrumental first step for designing personalized intervention delivery strategies. However, further research is needed to also identify barriers to dietary adherence. Identifying both individual and universal barriers will have important implications for exactly how to promote adherence, and allow for an even more targeted and personalized intervention approach.