The present study applied G4G 2.0 strategies, including stoplight color labeling, behavioral design, and an education and marketing campaign.¹⁶ Although G4G 2.0 introduced menu revisions, the current study only tested behavioral design strategies and did not modify existing or include new recipes.

From May–July 2017, over 200 foods and beverages were coded according to G4G 2.0 nutrition standards. Foods and beverage nutrition information were processed by the G4G’s web-based algorithm, which assigns items into Red: Eat Rarely, Yellow: Eat Occasionally, or Green: Eat Often categories.¹⁷ In cases where saturated fat content was missing from the recipe printouts, the web-based algorithm could not be used, and the appropriate color code was determined by DFAC dietitians.

CDC researchers and DFAC dietitians selected eleven strategies from the G4G 2.0 based on feasibility of implementation (Table 2). Researchers and DFAC dietitians created daily menu and performance plate signs and color-coded tags. For example, lunch menus featured green-coded items on the on the menu corners, and as possible, a less expensive, green-coded item next to more expensive yellow- or red-coded items. Communication materials, such as posters and table tents, were printed from the Army Public Health Center’s Health Information Products e-Catalog.¹⁸ Fifteen different green-coded “performance plate” specials were created from existing menu items, one for each weekday during the 3-week menu cycle. Structural improvements included fruit baskets and bottled water at the checkout, installation of no- or low-calorie sparkling water machines, and a new beverage cooler wrap depicting healthier vending items. Placement strategies included moving green-coded entrees and vegetables to the first positions in the main hot line, placing a green-coded vegetable next to the French fries at the grill station, placing green-coded items at the beginning of the salad bar line, and moving green-coded refrigerated self-serve items (eg, bottled water, yogurt) to eye-level and front positions and red-coded items (eg, pies, sugar-sweetened beverages) to less visible positions.

Intervention strategies were designed to be implemented by DFAC staff as a part of their regularly scheduled work. DFAC staff training was led by dietitians prior to the intervention start date. A procedure manual outlining strategies was developed to standardize practices across DFAC staff. DFAC staff implemented the intervention during lunch services from August 21, 2017, to December 22, 2017. Due to miscommunication between researchers and cafeteria staff, the fruit basket intervention was prematurely implemented during the fourth week of the baseline data collection period. Regular check-ins with CDC researchers, DFAC dietitians, and the food service manager served to address issues associated with program implementation. In the post-intervention period, staff were no longer trained or reminded to implement the intervention strategies but were not specifically instructed to cease implementing them.

Foods sales, the primary outcome, was assessed weekly from point-of-sales (POS) reports (MICROS Systems Inc, Oracle Corporation, Redwood Shores, CA). Intervention baseline POS reports were collected for the 18 weeks (6 menu cycles) from April to August 2017, implementation reports were collected for the 18 weeks from late August to December 2017, and post-intervention reports were collected for the 9 weeks (3 menu cycles) from late December 2017 to March 2018. Due to unintentional early implementation of the fruit basket strategy, baseline data period was only 3 weeks, the intervention period was 33 weeks, and the post-intervention period was 9 weeks. Mean sales volume for 16 healthy and less healthy food items that were expected to be impacted by the intervention were abstracted from sales data. These items included entrees (red, yellow, and green), hot vegetable and starch side dishes (grains, potatoes), grill station items (hamburgers/hotdogs, grilled chicken, vegetable/turkey burgers, and French fries), whole fruit, deserts (cakes, cookies, pies, and brownies), and self-serve refrigerated items (pudding, fruit cups, yogurt, hummus, and pre-packaged salads).

Data analysis was completed using SAS 9.4 and Stata 15.1 on weekly sales data (SAS Institute, Cary, NC) (StataCorp). Weekly food sales were calculated during the baseline, intervention, and post-intervention periods. Sales for each item were standardized by total sales volume to account for differences in overall sales over the study period. For each food outcome, the proportion of each food sold out of total foods sold during each menu cycle was multiplied by the average total number of foods sold per menu cycle. For example: (Actual # of green entreés sold in intervention week 1 / total # of all items sold in menu intervention week 1) * (overall total # of all items sold during entire study / 45 weeks total study duration). We compared mean weekly sales of each food outcome (adjusted for total weekly sales of all foods) between the baseline and intervention period and between the intervention and post-intervention period using t-tests.

Next, we estimated two negative binomial models for each food item. Model 1 was estimated over the baseline and intervention periods only (weeks 1–36) and included the following covariates: a two-way interaction between the intervention period indicator (=0 during baseline period, = 1 during intervention period) and continuous time trend (including main effects), total sales, the menu cycle week (1, 2, 3), and the special meals indicator (=1 if special meal occurred that week, = 0 otherwise). The immediate change from baseline to intervention was presented as the incidence rate ratio (IRR) and calculated as the exponent of beta-coefficient of intervention period indicator. The slope change from baseline to intervention period was presented as the IRR and was calculated as the exponent of beta-coefficient of the interaction effect between time trend and intervention period indicator. Model 2 was estimated over the intervention and post-intervention periods only (weeks 4–45 for whole fruit, weeks 19–45 for all other items) and included the following covariates: a two-way interaction between the post-intervention period indicator (=0 during intervention period, = 1 during post-intervention period) and continuous time trend (including main effects), total sales, the menu cycle, and the special meals indicator. The immediate change from intervention to post-intervention was presented as the incidence rate ratio (IRR) and calculated as the exponent of beta-coefficient of post-intervention period indicator. The slope change from intervention to post-intervention was presented as the IRR and was calculated as the exponent of beta-coefficient of the interaction effect between time trend and post-intervention period indicator.

Intervention implementation was assessed each menu cycle (3 weeks) using an environmental scan adapted from the Healthy Hospital Cafeteria Scan¹⁹ and the Military Nutrition Environmental Assessment Tool (m-NEAT).²⁰ Study fidelity measures how well intervention strategies were delivered according to plan.²¹ The modified scan was piloted by CDC researchers and DFAC staff and refined prior to the intervention. The environmental scan was completed during weekday lunch service three times during baseline, six times during the intervention, and twice during post-intervention. The same CDC researcher collected all scans to prevent inter-rater variability. We assessed the presence and placement of promoted ‘green’ foods/drinks, demoted ‘red’ foods/drinks, and communications materials (eg, posters, table tents). The scan also assessed the availability of color-coded tags correctly identifying items. Questions were binary (eg, yes/no), ordinal (eg, ≥5, 3–4, 1–2), or Likert-style (eg, Very Good, Good, Fair, Poor, Very Poor) for ease of data analysis. There was also an area to provide additional commentary on contextual factors (eg., staffing issues, quality of items). Scans were completed using a paper version and then entered in a web-based form built for the project (Epi Info, CDC, Atlanta, GA). Photographs of strategies were routinely taken to supplement fidelity data.

Fidelity statistics (percentages) were calculated for baseline, intervention, and post-intervention periods, and was determined for availability, prominent placement, and presence and correct placement of color-coded labels for green-coded hot foods, grab n’ go vended items, drinks, chips, salad bar, and menu boards. Prominent placement was defined as placing the promoted items either first in line, near the register, or at eye-level. We totaled values for each of the scales (eg, Yes/No questions were coded as 1 or 0, while a 5-point Likert scale were assigned values of 4 to 0) and divided by the number of assessments to create a fidelity average. We then divided the average by the highest value (eg, 4 for a 5-point Likert Scale) to create a fidelity percentage. For example, in the ‘Salad Bar’ category, the average score for ‘prominent placement’ across assessments was 3.17 out of a possible 4, thus, the fidelity percentage was 79% (3.17/4 × 100). High fidelity was defined as having a mean percentage of 75% across all categories, as defined in prior intervention studies.^22,23 Moderate fidelity was defined as 50%–74% and low was 0%–49%.²³