The design for this study includes two phases to examine the maintenance of a dietary behavior change in adults. Figure 1 provides an overview of the study flow and measurement time points. In the first phase (acquisition), an intervention was used to produce a change in F&V consumption in order to study, in a second phase (maintenance), if an intervention targeting maintenance could sustain that change. All interventions were delivered using a fully automated telephony system that used an interactive voice response (IVR) system to generate speech that emulated counseling by a trained behavioral counselor combined with a speech recognition system to understand what the participant said. To perform the study, a sample of healthy adults who consumed less than the recommended level of F&V (ie, ≤ 5 servings/day) were recruited. During the acquisition phase, they were given a tested dietary telephony intervention for up to 6 months, called Telephone-Linked Care (TLC)-EAT, which has been shown in previous studies to have positive effects on initiating dietary improvements, including increased F&V consumption [32,33]. Because both prior studies found an increase of approximately one serving of F&V for the TLC-EAT intervention group, it was hypothesized that this effect would occur in a new sample of participants. F&V intake was examined at enrollment and after 3 months of the acquisition intervention. Participants who increased their consumption of F&V by at least one serving/day became eligible for the second phase (maintenance). Participants who did not show a one-serving increase were allowed to participate for an additional 3 months only. Those who achieved a one-serving increase during the acquisition study were eligible to participate in the second phase (maintenance, randomized controlled trial). Participants who entered the second (maintenance) phase of the study were blinded to the inclusion criterion (eg, an increase of >1 serving/day of F&V). Participants in the maintenance phase were randomized to one of three groups, receiving one of two theoretically based TLC interventions targeting maintenance of behavior (TLC-Maintenance GST or TLC-Maintenance SCT) or to the control group (an assessment only without an intervention). All participants in the randomized controlled trial (RCT) were assessed every 6 months for a total of 2 years post randomization (baseline). The study protocol received full board review and approval by the Boston University School of Medicine Institutional Review Board (NCT00148525). See Multimedia Appendix 1 for the CONSORT-EHEALTH checklist [34].

The pool of possible participants was randomly selected from the voter registration lists from the Boston metropolitan area. Participants were required to be ≥18 years of age, live in the Boston area, have access to a touch-tone telephone, and be generally healthy. Participants were also required to be “under-consumers” of F&V as defined by eating fewer than 5 servings/day. However, in the first 2 months of recruitment, a large proportion (62%) reported consuming more than 5 servings/day using the National Cancer Institute F&V Screener (NCI FVS) [35]. Given that this measure may overestimate the servings of F&V [36-38], the inclusion criterion was modified to “consuming less than eight F&V servings/day” in order to correct for the bias in the NCI FVS. This modification decreased the percentage of screened individuals who were not eligible (39%) yet still captured adults who were not meeting a public health goal level, for example, the 2005 dietary guidelines of 9 servings/day of F&V [39-41]. Participants were not eligible if they were pregnant or if they had a recent health event such as a diagnosis of cancer, myocardial infarction, kidney disease, eating disorder, or were prescribed a special diet.

The University of Rhode Island Survey Research Center performed study recruitment, assessment, and randomization. The Center randomly selected names from a voter registration list and mailed a letter inviting them to participate in the study. The letter was followed by a phone call from the Center about one week later where the interviewer described the study, screened for eligibility (see Participant section above for eligibility criteria), obtained verbal consent, and administered study instruments. Research assistants at Boston University School of Medicine, where the intervention (ie, TLC-EAT) was hosted, sent participants an enrollment packet in the mail, which contained a welcome letter, consent form for their records, and a manual and personal password for the healthy eating acquisition intervention (ie, TLC-EAT). A week after enrollment, these research assistants called participants to train them on how to use TLC-EAT. Participants did a practice call with the research assistant on the line. Those who completed the training were transferred to the first intervention call (the “training call”) with TLC-EAT. Thereafter, all TLC calls were outbound calls initiated by the automated system, which called participants at the time initially entered into the automated scheduling system during the training call. Participants were asked to complete one TLC-EAT call per week for 12 weeks with the option of rescheduling any of the incoming TLC calls or initiating calls to TLC if preferred. After 12 weeks, F&V consumption was reassessed with the NCI FVS to determine eligibility for the maintenance phase of the trial. Participants who increased by one serving of fruit and/or vegetable were eligible to be randomized to one of three groups for the RCT. Random allocation to group assignment was generated by the SRC’s computer program that used urn randomization protocols to balance groups by gender. Those who did not increase by one serving after 3 months were invited to continue with the acquisition intervention, TLC-EAT, for an additional 3 months and then were assessed again for eligibility into the RCT. Those who qualified and agreed to continue in the RCT were assessed at baseline (randomization), 6, 12, 18 and 24 months post baseline. A random sample of those who qualified for the RCT were invited to complete a blood draw at the university’s General Clinical Research Center at baseline only. Assessments during the RCT included the NCI FVS, 24-hour dietary recalls, and psychosocial measures. Participants received US $20 for completing a survey at baseline, 6-month, 12-month, 18-month, and 24-month time points, and $50 for the in-person, fasting blood draw.

Sample size for the RCT was based on 80% power to detect a 20% difference in the percentage maintaining an improvement in F&V consumption at 24 months. This sample size also provides 80% power to detect a small-to-medium effect size for differences in mean F&V consumption based on Cohen’s definition [57]. The primary hypothesis was that the treatment groups would maintain their initial gains in F&V consumption better than the control group, and the TLC-GST group would be significantly more successful in achieving this outcome than the TLC-SCT group (ie, GST > SCT > Control). This was tested defining maintenance as sustaining a one-serving increase in F&V that was achieved during the acquisition phase (as this is the eligibility criterion for the RCT). This primary hypothesis was tested using two approaches: a categorical approach and a continuous approach. The categorical hypothesis tests the proportion of the group who maintain, while the continuous tests the differences in F&V serving size. For the categorical hypothesis, the sample size estimation was based on a 20% difference in the proportion of participants meeting the definition of maintenance between groups. If the proportion maintaining gains is 80% in one group, 60% in another, and 40% in a third, there is 80% power of detecting pairwise differences between groups. Expecting 5% attrition at every 6 months from baseline to 24-month time point, a sample size of 405 (n=135 per group) is required at baseline and will yield a final sample of 330 (n=110 per group). This sample size provides 81% power of detecting the difference at 24-month time point between TLC-SCT and Control Groups (60% vs 40%), 87% power of detecting the difference between TLC-GST and TLC-SCT Groups (80% vs 60%), and 99% power of detecting the difference between TLC-GST and Control Groups (80% vs 40%), testing at the two-tailed alpha .05 level. For the continuous hypothesis, group differences are measured by the mean change from baseline to 24-month time point. A sample size of 110 provides 80% power of detecting an effect size of .38 of the standard deviation of the change score in F&V servings.

Data will be analyzed using SAS, version 9.1 for Windows. Alpha level was set at P<.05. Descriptive statistics will be used to characterize the different study subject samples before and after the acquisition study period. Descriptive statistics will examine those individuals who were screened for eligibility and those who qualified and enrolled into the RCT. Sample comparisons between those who were eligible and those who were not will be made using data collected at enrollment (screening data). Comparisons of the different study samples will be performed using Student’s t test for continuous variables and Pearson’s chi-square test for categorical variables. An additional analysis will be performed evaluating the change in F&V servings from initial screening (enrollment) to baseline (ie, the beginning of the maintenance phase) for the sample who was randomized into the RCT. An intention-to-treat approach using the last observation carried forward approach will be used to include those participants who drop out of the study or for whom there are missing data.

The primary hypotheses for the RCT is that the treatment groups will maintain their initial gains (ie, one-serving increase) in F&V consumption better than the control group, and the TLC-GST group will be significantly more successful in achieving this outcome than the TLC-SCT group (ie, GST > SCT > Control) on both a categorical (primary hypothesis) and a continuous measure (secondary hypothesis) of maintenance at all follow-up time points. The primary time points for the study will be baseline (randomization), 6 months (post intervention period), and 24 months (end of the follow-up period). The primary comparisons will be considered from baseline to 6 months, and baseline to 24 months. Group differences will be compared using a continuous variable (F&V consumption) at these time points. We will also compare groups on the percent who maintained a one-serving/day or greater increase in F&V consumption. The categorical measure of maintenance will be analyzed through multiple logistic regression models, and a continuous measure of maintenance will be analyzed through multiple linear regression models. Independent variables in these models will include a set of indicator variables for study group, baseline levels relating to each outcome variable, and potential confounders (variables found to differ between groups at baseline).

Longitudinal models will be used to explore group differences on dietary indicators over time, using data from all study evaluation points. These analyses will use the Generalized Estimating Equations (GEE) approach to accounting for the longitudinal nature of the data by modeling the within-subject correlation and adjusting both regression parameters and standard errors for this correlation. As compared to traditional repeated measures analysis of variance, the GEE approach allows for the inclusion of all available data from subjects with incomplete follow-up in the analysis. For categorical outcome measures, GEE logistic regression models for longitudinal data will be used, while for continuous outcome measures, GEE linear regression models will be used. Independent variables in these models will include a set of indicator variables for group, a set of indicator variables representing time (with baseline taken as the reference group), and a set of interaction terms modeling differential changes over time for the three study groups. Potential confounders identified in preliminary analyses will also be included in these models. Our primary interest is in the interaction terms, which will test whether the pattern of change in consumption of F&V over time differs by group.

While the primary focus of our analyses of the number of F&V servings is on changes from baseline to 6 to 24 months, our longitudinal models will also allow us to examine maintenance decay in the number of servings from 6 months (at the end of the maintenance intervention) to 24 months. Similarly, our longitudinal models for the categorized maintenance outcome will allow us to examine changes in maintenance from 6 to 24 months as well.

Secondary analyses will examine the influence of psychosocial variables. For TLC-SCT, we hypothesize that the outcomes are mediated by self-efficacy. For TLC-GST, we hypothesize that outcomes are at least partially explained by changes in goal system variables such as levels of inter-goal facilitation, inter-goal substitution, and inter-goal conflict at all major follow-up time points.

These hypotheses will be tested with path analysis. Potential mediational pathways of the effect between the randomized groups and F&V intake for the theoretical construct variables will be examined. Path models will be constructed to test the direct and indirect associations indicated by our research model. Using 6 and 12 month data as indicators of processes of maintenance in both the mid-term (at the end of the maintenance intervention) and in the longer term (6+ months post intervention), path analyses, both separate and combined, will be conducted to examine mediators of maintenance variables as influenced by group assignment. To examine the TLC-SCT change model, paths will be modeled from an intervention variable to self-efficacy to F&V intake. Direct and indirect effects of the intervention will be estimated through standardized path coefficients. For the TLC-GST, goal system variables will be examined in analogous path models.

Cost analyses are planned for the study conditional upon demonstrating that the maintenance interventions are effective in altering and sustaining improvements in diet. The analysis will be based on the recommendations of the US Public Health Service Panel on Cost-Effectiveness Analysis [58]. An incremental cost-effectiveness ratio on the acquisition phase will be computed. The incremental cost-effectiveness of the two maintenance intervention conditions will be compared relative to the acquisition intervention to assess the resource use associated with incremental sustained improvements in health.

There are a number of important issues to consider about this study. The overarching goal of this study is to better understand how to help individuals maintain a newly acquired health behavior. The challenge was first to identify a population that is engaging in unhealthful behaviors and that the behavior is amenable to change. F&V was selected as the principal target for the intervention study for multiple reasons. First, the majority of the US population is not consuming enough F&V, providing an opportunity to recruit a large enough sample within a reasonable time frame. Second, we had to identify an evidence-based, off-the-shelf intervention that would likely produce a change in F&V in order to study whether a newly acquired behavior can be sustained with interventions specific to the maintenance of that newly acquired behavior. While the TLC-EAT intervention was used for the acquisition phase of the study, we had to consider how much time it would take to achieve a primary intervention effect using TLC-EAT to make a meaningful positive dietary behavior change to qualify sufficient numbers of participants for the maintenance intervention RCT, and what should be considered a positive response in this RCT. One option was to offer TLC-EAT acquisition intervention for as long as it took the participant to show a positive improvement. This was not feasible given the time limitations of the study but fortunately was not a factor in the study design because the TLC-EAT acquisition intervention was known to achieve a positive effect within 3 months based on a previous study [31]. We also decided to define a positive intervention effect in the acquisition phase as an increase in F&V consumption by one or more servings per day. We did consider using a threshold criterion of a successful acquisition intervention effect of 5 servings a day but did not want to exclude individuals who might begin the acquisition phase at very low levels of F&V consumption and have a substantial acquisition intervention effect but still not meet the threshold of 5 servings of F&V/day. Moreover, including individuals who increase at least one F&V serving a day is in line with the usual level of change achieved in successful dietary change programs [6]. In addition, there is no evidence in the literature that increasing F&V consumption from 0-1 serving/day confers different risk reduction than going from 4-5 servings/day. Moreover, it is not known if it is easier to maintain a one-serving increase from 4-5 servings per day compared to a change from 0-1 servings per day, or whether a larger increase in F&V consumption (for example 2 or more servings/day) is more difficult to maintain than a smaller increase during behavior change acquisition.

A major consideration for our study was how to define maintenance. There is not a well-accepted definition of maintenance of a behavior change. It could be defined as a criterion outcome, which means that the individual consumes at least the same number of servings of F&V at each outcome measurement point as he or she consumed at the end of the acquisition phase. Another approach is to compare the three groups at each time point using the proportion of participants in each study group who reach criterion at that time (eg, consume at least the same number of servings of F&V at the end of acquisition to the end of the maintenance assessment). In doing so, we will consider the standard error of the measurement of F&V intake so that maintenance of F&V consumption will be defined as at least the same intake the person achieved at the end of the TLC-EAT acquisition period minus the standard error. Another approach is to compare changes in intake of F&V between randomization and each outcome measurement point. These “change scores” can be compared across the three study groups. The advantage of this approach over the criterion outcome approach is that it considers both the degree of preservation of intervention effect and possible increases in intervention effect by the maintenance intervention. Recent research suggests beneficial health effects as F&V consumption increases rather than only among those who meet a certain threshold (ie, 5 servings/day) [2]. The advantage of using a “change score” rather than a “threshold score” is that it may be the most relevant public health criterion. However, it is a less intuitive measure of maintenance than the criterion measure. Thus, both approaches will be examined.

The last consideration concerns the diet content of the basic TLC-EAT intervention. TLC-EAT has shown to improve F&V in two previous interventions yet the topics covered in the intervention calls include other food groups (diary, protein, grains). The approach taken in TLC-EAT relies heavily on the strategy of substitution of “healthy” for “unhealthy” foods; in doing so, it promotes greater consumption of some food groups, like F&V, possibly at the expense of others, like red and processed meats, and recognizes the public health benefits of targeting multiple dietary risk factors and the practical realities of intervening on diet. Thus, the diet topics in all of the interventions were written using this approach.

While few studies have evaluated maintenance-specific interventions for dietary change, fewer maintenance-specific interventions have been studied in which outcomes exceed 12 months, and even fewer have compared theory-guided interventions. The major innovation of this project is the design, creation, and testing of a maintenance-specific intervention (TLC-GST) and its comparison with an intervention based on the most commonly used behavioral theory, SCT. GST is a “maintenance-focused” theory in that it posits that many worthy and attainable goals fall to the wayside because other endeavors, for any number of reasons, take resources away from their pursuit. If it is competition for internal and external resources that triggers a significant proportion of relapse, then an intervention that successfully advocates for the continued allocation of resources to dietary behavior, as well as assists users in more efficiently managing the resources they have across their goals, should make a difference. It is also possible that maintenance of goal behaviors that do not have immediate and perceivable rewards, especially true of long-term risk reduction, is especially appropriate for an intervention that helps balance demands from goals that have more short-term and palpable rewards.

Few studies have compared theories, mostly because it is a challenging task that requires that at least two interventions be available that are the same or very similar with each of them based on a different behavioral theory. There are very few pairs of behavioral interventions that have functional comparability while being driven by distinctly different theories. Practically speaking, a researcher has to design the two interventions being studied, and this is time-consuming and expensive. Nonetheless, this is exactly what we did in creating two theoretically guided interventions for maintenance of dietary behavior change. Using SCT to map out a maintenance intervention was doable because it is one of the most widely used theories for changing and maintaining health behaviors. The relapse prevention model [53], which is informed by SCT, provided a practical framework for implementing a SCT-based behavior maintenance intervention by explicitly identifying which intervention strategies to include (eg, goal setting, planning ahead) to maintain a behavior change. The resulting TLC-Maintenance SCT intervention has good fidelity to SCT. In contrast, and to our knowledge, GST has never been applied to health behavior change, whereas it has a strong theoretical and experimental underpinning for application to health behavior and its change. No study has attempted to translate and apply GST for practical dietary interventions delivered to individuals over the telephone.

Another strength of the study design in our RCT is our ability to control for fidelity in delivering the interventions. The acquisition intervention, TLC-EAT, was delivered to all enrolled participants in the same way because it was delivered using an automated computer system. TLC-EAT is designed to provide a consistent intervention of all users that is tailored to the participant. TLC-Maintenance SCT also provides a consistent maintenance intervention for all of its users. Although a computer-controlled health behavior change intervention may or may not be as tailored as a human counselor intervention, the intervention content is explicit and discoverable and has complete fidelity. If differences are found between the two health behavior change maintenance interventions, they will be due to the underlying content of the interventions and not due to differences in provider training, provider skill level, or lack of adherence to protocol [59].

This study used a novel study design to initiate and then promote the maintenance of dietary behavior change through the use of an evidence-based fully automated telephony intervention. After the first 6 months (the acquisition phase), we will examine whether two telephony interventions built using different underlying behavioral theories, were more successful than an assessment-only control group in helping participants maintain their newly acquired health behavior change.