We developed a survey for restaurant food workers based on the constructs proposed by Griffith et al. (2010a) and previously administered surveys (Ball et al., 2010; De Boeck et al., 2015; Neal et al., 2012). A workgroup composed of EHS-Net health department staff designed the survey to apply to all types of restaurants, rather than for a specific company as previous researchers have done (i.e., our survey assessed handwashing resources, something that is relevant in all restaurants). The survey asked food workers to self-report their level of agreement with 28 statements (Table 2) using a Likert-scale ranging from 1 (strongly disagree) to 5 (strongly agree). Four of these items were reverse-coded (DeVellis, 2003). Table 2 includes the survey items, the number of responses, the mean scores (higher scores show stronger agreement [or disagreement for reverse-coded items]), and standard deviations. We also included questions to assess food workers’ food safety knowledge and experience working in restaurant kitchens.

To increase restaurant participation in the study and food worker honesty in the survey responses, study data collection was anonymous. Thus, to ensure we did not collect data that could allow the identification of food workers, we asked limited questions about individual demographics. For example, we did not collect data on staff race, ethnicity, or age.

The instrument was pilot tested with three restaurant food workers for comprehension and length of time to complete the survey. The survey was translated into Spanish by one native speaker and translated back into English by another native speaker to verify the translation. A copy of the food worker survey is provided in the Supplementary material, and all the study materials are posted at https://www.cdc.gov/nceh/ehs/ehsnet/study_tools/index.htm.

The study sample consisted of randomly selected restaurants in each of the eight EHS-Net health department’s jurisdictions. Restaurants were defined as establishments that prepare and serve food or beverages to customers but are not institutions, food carts, mobile food units, temporary food stands, supermarkets, restaurants in supermarkets, or caterers. In each EHS-Net jurisdiction, staff chose a geographic area in which to recruit restaurants for study participation, based on a reasonable travel distance (mean=88.1 min, range = 30 min to 4 h). One jurisdiction was urban; the other seven were a combination of urban, suburban, and rural areas. The staff then sent a list of restaurants within that area to CDC, which selected a random sample of restaurants for each jurisdiction. Staff in each jurisdiction requested voluntary study participation from managers in a random sample of restaurants and scheduled a data collection visit through telephone calls or visits to the participating restaurants. Within each restaurant, food workers were requested to voluntarily participate in the study. No incentives were provided to participate in this study.

Data collection took place from March 2018 to March 2019. Data were collected by EHS-Net staff. All data collectors participated in training designed to promote data collection consistency. At each restaurant, data collectors, EHS-Net staff, interviewed a manager (someone who had authority over the restaurant) about restaurant characteristics, asked food workers (staff members who prepare food in the restaurant) to complete a survey, and conducted an observation of food preparation and storage practices in the kitchen area. This paper presents data from the food worker survey on food workers’ beliefs about food safety.

Food workers completed a self-administered survey about their beliefs around food safety in their restaurant. The survey was provided in English and Spanish using the SurveyMonkey (Momentive, San Mateo, CA, USA) online survey platform. Food workers could either complete the survey online using the online application (at their convenience) or complete a paper version of the survey during the data collection visit. Any surveys completed on paper forms were entered into SurveyMonkey later by the data collectors. We did not record whether a food worker completed an electronic or paper-based survey.

A study identifier was used to link worker survey data to the matching restaurant; however, we did not collect data that could identify individual restaurants, managers, or workers. Each EHS-Net jurisdiction’s institutional review board approved the study protocol.

We randomly split the completed survey responses into two groups (n = 248 per group) for analysis. One group was used for model building, and the other group was used for validation of the statistical model. We examined the model fit of the theoretical model of food safety culture based on the constructs proposed by Griffith et al. (2010a) using confirmatory factor analysis. To assess fit, we examined the overall concordance of multiple indices; these included the chi-square (not statistically significant indicated a better fit), the standardized root mean square residual (SRMR ≤ 0.08 indicates a better fit), the comparative fit index (CFI ≥ 0.95 indicates a better fit), and root mean square error of approximation (RMSEA < 0.06 indicates a better fit) (Schreiber et al., 2006). However, the data that we collected did not support this model structure. Therefore, we then conducted an exploratory factor analysis to identify the factors that were empirically supported. We retained items that loaded onto unique common factors, had a primary factor loading of 0.4 or above, and did not load onto another factor at 0.3 or above. We then examined whether the data would benefit from a data reduction method using Bartlett’s test of sphericity, where a significant value supports further data reduction. We then examined the number of factors that would be supported by the model using a scree test and minimum average partials test (Velicer et al., 2000). Once we identified the four factors and their associated items, we assessed scale reliability using Cronbach’s alpha with an alpha coefficient of 0.65 or higher considered acceptable. After that, we conducted structural equation modeling to identify an appropriate model structure and to determine if the data would support further generalization to a higher-order factor. Finally, we created a composite measure for each factor in the model, based on the structural equation model, where the sum of the Likert-scaled questions for each factor was calculated. Negatively phrased questions were recoded so that higher scores would equate to positive agreement (e.g., strong disagreement with a negatively phrased item was recoded as strong agreement for analysis). We then divided the sum by the number of questions associated with the factor to provide a standardized score for each factor. We used SAS version 9.4 (SAS Institute, Cary, NC, USA) to analyze the data.

We contacted 1,496 restaurants to participate in the study, of which 506 were excluded (restaurants were no longer in business, were not a restaurant [e.g., a grocery store], the manager was unable to communicate with the study recruiting staff in English). Of the remaining 990 restaurants, we had participation from 331 restaurants. We received 579 food worker survey responses from those 331 different restaurants. Manager interview data indicated that the study restaurants were largely independently owned (57.1%). Food worker survey data showed that the largest group of food workers had 1–5 years of experience (39.9%), 1–5 years of tenure in their current establishment (46.6%), and worked primarily in the kitchen (55.4%). Most respondents had a current Certified Food Protection Manager credential (56.7%); however, only 9.5% were in a supervisory role. Food workers reported primarily speaking English (72.5%) and Spanish (18%). Most of the food workers had completed high school (32.5%), had at least some college education (49.1%), and were male (50.6%) (Table 1).

Of the 579 responses, 44 (7.6%) were completed in Spanish. All Likert-scaled items were initially tested for multicollinearity, deviation from linearity, consistency with similar items, and if all items were answered. This screening led us to drop item 5 (Table 2) from further analyses because it did not consistently correlate with other similar questions. This lack of correlation is likely due to the influence of varied glove use requirements across the jurisdictions.

An initial theoretical model based on Griffith et al. (2010a) work was constructed where we associated each of the items to one of five constructs: Commitment; Communication; Leadership; Resources; and Risk Awareness. We then assessed this model for fit. However, our data did not support this model. None of the fit indices — χ²(485) = 2,039.45, p < 0.0001; SRMR = 0.10; CFI = 0.68; and RMSEA = 0.11 (0.12, 0.11) — indicated an adequate fit between the data and the model.

We initially included 27 items in a factor analysis. Sixteen questions were retained in the model. The principal factor analysis used squared multiple correlations with all other items, unweighted least squares factors, and a promax (oblique) rotation. The remaining 11 items did not load onto common factors or meet the above criteria for retention in further analyses.

A significant Bartlett’s test of sphericity (χ²[62] = 133.92, p < 0.0001) indicated that the data could be reduced into factors. Results of a scree test and a minimum average partials test suggested four factors would be sufficient to explain the variance (Velicer et al., 2000).

Table 3 shows the survey items and factor loadings. The EHS-Net working group reviewed the items that formed each of the four factors to provide their perceptions of the constructs measured by each factor. The working group labeled those factors as Leadership, Management Commitment, Employee Commitment, and Resource Availability. Leadership included six items, Employee Commitment included four items, and Resource Availability and Management Commitment included three items each.

Scale reliability for each factor was assessed using Cronbach’s alpha; each factor had acceptable reliability (Leadership = 0.88, Employee Commitment = 0.87, Resource Availability = 0.72, Management Commitment = 0.73) (Nunnally, 1978). External validity was assessed using the reserved half of the dataset; the results were similar to those obtained from the first half of the data.

Structural equation modeling was used to identify the relationships among the factors and to determine if the data would support a higher-order factor (Anderson and Gerbing, 1988). In other words, this modeling was to determine if the identified factors are stand-alone factors, are inter-related, and if they can be further generalized to a higher-order factor (an overarching factor that is explained by these primary factors, similar to how individual questions explain the primary factors).

We examined various structural forms of the factors identified in the exploratory factor analysis; a model with one higher-order factor (Worker beliefs about food safety culture) was found to be optimal. The fit indices for this model showed an overall good fit — χ²(100) = 210.78, p < 0.0001; SRMR = 0.05; CFI = 0.95; and RMSEA = 0.07 (0.08,0.05) (Figure 1).

Reliability estimates were generally acceptable (Table 4). Item reliability was generally above 0.5, except for item 7 (0.27). We chose to retain this item because of its contextual similarity to other items and to maintain factor reliability. The four primary factors exhibited acceptable overall composite reliability (Leadership = 0.91, Employee Commitment = 0.89, Resource Availability = 0.79, and Management Commitment = 0.78). Relationships between individual items and their associated factor were examined; all pathways were significant. Similarly, the relationships between each of the primary factors were significantly associated with the higher-order factor (Table 4). The finding that the t-values are significant for these path coefficients suggests that the items are measuring the same construct.

All constructs had composite scores spanning the entire range, from 1 (strongly disagree) to 5 (strongly agree). In general, food workers viewed each of the factors positively (composite score >3), although individual workers in some restaurants reported lower scores. Food workers viewed Resource Availability highest (mean = 4.69, SD = 0.57), followed by Employee Commitment (mean = 4.49, SD = 0.62), Leadership (mean = 4.28, SD = 0.69), and Management Commitment (mean = 3.94, SD = 1.05). The overall belief in food safety culture had a mean score of 4.35 (SD = 0.53) (Table 4).