Study design and setting
To understand individuals’ responses to the early outbreak of COVID-19, we asked study participants how likely it was that the spread of the coronavirus will cause a national public health crisis and their level of concern about contracting COVID-19 by attending campus events. Data were collected from participants in an ongoing study focused on understanding consumer food choice and consumption behavior during midday meals (11:00 AM – 2:00 PM). Students, staff, and faculty of Louisiana State University (LSU) were recruited to participate in a study held at the Food Sensory Services Lab on the Baton Rouge campus where they were offered a choice among several commercially prepared lunch options. They were provided a fixed budget for lunch and kept unspent budget as cash compensation. After providing informed consent, participants moved to isolated, individual kiosks with a computer to answer an online survey in which information treatments were randomly assigned and participants chose among a series of competing lunch options. One of the participant’s preferred lunch options was delivered by staff to the kiosk. Upon completing the meal, staff removed the food tray and the participant completed an online exit survey via the kiosk computer that focused on satisfaction with the provided meal and personal information.
Randomly assigned experimental elements included whether participants received information about food waste (vs. screen time, Food Waste Info); received information about improving nutrition (vs. financial literacy, Nutrition Info); received meals with more vegetables (vs. fewer, Vegetable Group); received meals on a large plate (vs. smaller, Large Plate); received meals on a compostable plate (vs. plastic, Compostable Plate); and received menus where the vegetable was listed at the top in the description of the offering (vs. lower, Veg Top of Menu). More detail and context concerning the experimental elements are included in the Supporting Information (Figure S2).
The food preference study initially began on February 17, 2020. In late February, as concerns about the spread of COVID-19 in the United States increased, we feared that the national emergence of COVID-19 could influence the composition of participants volunteering for this in-person study. To control for such potential changes, we added two questions to all exit surveys administered from March 3 to the final day of the study on March 12, 2020 (Figure S1): (1) In your opinion, how likely is it that the spread of COVID-19 (the coronavirus) will cause a public health crisis in the United States? (National Likelihood); and (2) How concerned are you that you will contract COVID-19 by attending events on campus (Local Vulnerability)? Responses to both questions were registered on a 5- point Likert scale (1 = very unlikely/unconcerned, 2 = moderately unlikely/unconcerned, 3 = neither likely/concerned nor unlikely/unconcerned, 4 = moderately likely/concerned, 5 = very likely/concerned). Responses to these two questions can provide insights into local perceptions of contemporaneously delivered state and federal COVID-19 communications critical for informing strategies to encourage public adherence to safety guidelines.
LSU continued all in-person classes and food service operations through March 13, 2020, and no official announcements were made regarding the cancellations of any on-campus activities before the end of the last study session (2: 00 PM March 12th) [11]. At 4:00 PM on March 12th, 2020, LSU’s official communications regarding COVID-19 first mentioned the cancellation of on-campus classes starting with the week of March 16th [12], and then announced the cancellation of non-class activities involving 30 people or more immediately at 11:30 AM on March 13, 2020 [13]. For reference, a national emergency was declared in response to COVID-19 the afternoon of March 13, 2020 [14].
Participant recruitment and sampling
The sample used in this study included 356 participants enrolled from March 3rd through March 12th, 2020. Individuals were recruited via pre-existing email recruitment lists, flyers circulated on campus, advertising announcements in classes, and advertisements in university locations. Inclusion criteria included age 18 years or older with no dietary restrictions involving beef consumption. Three participants were omitted from all analyses because they failed to pass an attention test embedded in the survey, leaving an effective sample size of 353.
Analysis
Descriptive statistics and regression analyses were conducted in Stata (version 16) and classification tree analyses were conducted using R (version 3.6.0). When more convenient for analysis, the 5-point Likert responses were simplified into binary variables (very or moderately likely/concerned = 1; all other responses = 0). We also defined the variable National, not Local to equal one when participants think a national crisis is very or moderately likely (National Likelihood > 3) but are neither very nor moderately concerned about contracting the virus by attending campus events (Local Vulnerability ≤ 3). Personal characteristics applied in the analyses included gender, age, student status (=1 if enrolled in University classes, =0 otherwise), household income, race, health insurance status, recycling frequency, experience with food composting, previous knowledge of food waste as an issue, whether they were trying to eat healthier, and whether they attended the session in response to in-person flyer distribution on the experiment date (as opposed to alternative recruitment such as emails or class announcements).
The timing of the study was included as a control variable in the model using several approaches. In one variant, we controlled for the number of confirmed cases nationwide in the United States as of 2:00 pm on the day of the study [15], while in another variant we controlled for the number of confirmed cases locally in Louisiana as of 2:00 pm [16,17,18,19,20]. A third variant, presented in the Supporting Information, controlled for the timing of the study with daily fixed effects (e.g., a separate variable denoting that the day of the study was March 3, March 4, etc.). In addition, personal characteristics were included in an ordered logit regression model of the Likert scale response to the two COVID-19 perception questions and in a binary logit regression model for the National, not Local variable. Statistical significance was set at the 5% level with results at the 10% level deemed marginally significant.
We also estimated classification trees for each dependent variable (the binary versions of National Likelihood and Local Vulnerability, plus National, not Local). Classification tree estimation is distinct from logit regressions in that they predict observations that are likely to be classified into a particular group (e.g., those for whom National Likelihood = 1) by splitting the sample into multiple sub-samples based on a series of predictor variables [21, 22]. The predictor variables included the explanatory variables used in the logit models. A version of classification tree analysis where the Gini improvement measure was used as the splitting criteria [23] was employed. A tree was grown first by splitting the sample based on the outcome of one of the predictor variables, and then branches were created to further subdivide the sample based on the levels of additional predictor variables. While classification tree analyses yield numerous outputs (see Supporting Information for standard graphical outputs), we focused on the relative variable importance scores, which measured the variable’s ability to predict the outcome in the estimated tree. Relative variable importance scores provided insights, for example, into which of several explanatory variables that are significant in a logit regression prove most critical for improving predictive capacity.