key: cord-0702563-dwuu0760
authors: Eshun-Wilson, Ingrid; Mody, Aaloke; McKay, Virginia; Hlatshwayo, Matifadza; Bradley, Cory; Thompson, Vetta; Glidden, David V.; Geng, Elvin H.
title: Public Preferences for Social Distancing Policy Measures to Mitigate the Spread of COVID-19 in Missouri
date: 2021-07-08
journal: JAMA Netw Open
DOI: 10.1001/jamanetworkopen.2021.16113
sha: 08db39c46e7aa0a7fc081e433ce421bf9e7bd377
doc_id: 702563
cord_uid: dwuu0760

IMPORTANCE: Policies to promote social distancing can minimize COVID-19 transmission but come with substantial social and economic costs. Quantifying relative preferences among the public for such practices can inform locally relevant policy prioritization and optimize uptake. OBJECTIVE: To evaluate relative utilities (ie, preferences) for COVID-19 pandemic social distancing strategies against the hypothetical risk of acquiring COVID-19 and anticipated income loss. DESIGN, SETTING, AND PARTICIPANTS: This survey study recruited individuals living in the Missouri area from May to June 2020 via randomly distributed unincentivized social media advertisements and local recruitment platforms for members of minority racial and ethnic groups. Participants answered 6 questions that asked them to choose between 2 hypothetical counties where business closures, social distancing policy duration, COVID-19 infection risk, and income loss varied. MAIN OUTCOMES AND MEASURES: Reweighted population-level relative preferences (utilities) for social distancing policies, subgroups, and latent classes. RESULTS: The survey had a 3% response rate (3045 of 90 320). Of the 2428 respondents who completed the survey, 1669 (75%) were 35 years and older, 1536 (69%) were women, and 1973 (89%) were White. After reweighting to match Missouri population demographic characteristics, the strongest preference was for the prohibition of large gatherings (mean preference, −1.43; 95% CI, −1.67 to −1.18), with relative indifference to the closure of social and lifestyle venues (mean preference, 0.05; 95% CI, −0.08 to 0.17). There were weak preferences to keep outdoor venues (mean preference, 0.50; 95% CI, 0.39 to 0.61) and schools (mean preference, 0.18; 95% CI, 0.05 to 0.30) open. Latent class analysis revealed 4 distinct preference phenotypes in the population: risk averse (48.9%), conflicted (22.5%), prosocial (14.9%), and back to normal (13.7%), with men twice as likely as women to belong to the back to normal group than the risk averse group (relative risk ratio, 2.19; 95% CI, 1.54 to 3.12). CONCLUSIONS AND RELEVANCE: In this survey study using a discrete choice experiment, public health policies that prohibited large gatherings, as well as those that closed social and lifestyle venues, appeared to be acceptable to the public. During policy implementation, these activities should be prioritized for first-phase closures. These findings suggest that policy messages that address preference heterogeneity (eg, focusing on specific preference subgroups or targeting men) could improve adherence to social distancing measures for COVID-19 and future pandemics.

This supplemental material has been provided by the authors to give readers additional information about their work.

Attribute selection:

We sought to identify attributes that were unconfounded, which is to say unlikely to be both representations of an underlying common but unsolicited preference, as well as present participants with a range of response categories that was wide enough to capture significance heterogeneity in preferences but within a range where a linear relationship was considered plausible for continuous attributes. We identified several candidate social distancing policy features of importance, including: (1) the duration of the policy, (2) the clarity of the messaging regarding the policy end date, (3) the closure of, childcare services, schools and colleges, indoor lifestyle services (e.g. salons, bars), outdoor recreation services (parks, beaches), religious services and mass gatherings. In addition, we determined that risk of infection or hospitalization for the individual and others, as well as income loss were other key determinants of adherence to social distancing public health measures. Through an iterative process of brainstorming/discussion, reducing and merging attributes to prevent overlapping concepts, reduction of number attributes to minimize cognitive burden, removal of inappropriate attributes and refinement of wording we refined attributes.

In the experiment design we sought to balance pragmatism and completeness and therefore limited the number of attributes according to DCE design guidelines (five to seven attributes) and selected those attributes which we determined to be key decision drivers and of the greatest public health policy significance during the time period. To further maximize statistical and response efficiency (avoid fatigue in respondents) we limited the number of attribute levels (<=3) and the number of prohibited attribute level combinations and limited the number of DCE questions asked of each respondent to six and opted for two policy scenarios per task. We manually removed combinations considered non sensical. The final design presented consumers with two potential counties, with different sets of policies, and sought to understand which location participants preferred, all else being equal. Each policy reflected 7 attributes related to the opening or closure of social venues, education facilities and outdoor activity services, whether large gathering were permitted, the duration of the policy, the potential income lost during the first six months after the policy was instituted and the associated underlying risk of COVID infection in the county (eTable 1).

To achieve statistical efficiency, we constructed a near balanced (i.e., each level appears equally often across the experiment) and near orthogonal (i.e., each pair of levels across attributes appears equally) design-based on a design of 7 attributes, 4 with 2 levels and 3 with 3 levels and 6 choice sets (questions) with two scenarios each. We additionally prohibited two attribute combinations in the design -"permitted large gatherings" and "risk of COVID infection -low" and "prohibited large gatherings" and "risk of COVID infection -high". We tested the design efficiency using the logit efficiency test in Sawtooth software with simulated data to obtain an efficient design with standard errors of 0.05 or less for the main-effects analysis for the estimated sample size of 600 participants.

Sample size estimation: We based our sample size calculation on the formula N ≥ (500 x c)/(a x t) -where N is the number of participants, t is the number of choice tasks (questions), a is the number of alternative scenarios and c is the largest number of attribute levels for any one attribute, and when considering two-way interactions, 'c' is equal to the largest product of levels for any two attributes -(500 x 9/ 2 x 6) (1). To additionally conduct subgroup analyses, at least 200 participants per subgroup is recommended. The DCE was powered to detect main effects and evaluate at least 3 subgroups (minimum calculated sample size of 600). We followed the Professional Society for Health Economics and Outcomes Research (ISPOR) guidelines for design of choice experiments (2, 3) .

The DCE was conducted in Missouri, a Mid-Western state in the US, with a population of 6,137,428. The majority of the population is white (83%) and 12% is Black/African American (4, 5) . We used randomly allocated social media advertising on Facebook and Instagram to recruit participants in the state. In addition, the survey was distributed via email to study volunteer networks, and to obtain preferences of Black/African We carried out one round of cognitive interviews and piloted the final survey questions iteratively to ensure intelligibility and coherency. The survey was programmed using Sawtooth Software and participants completed the survey using personal mobile devices or computers. Participants were randomly allocated to one of 300 versions of the choice experiment and the order of the attributes within each question was randomized.

Analysis:

Choice experiment modelling is based on random utility theory (RUT) which assumes that the utility (U) for individual i conditional on choice j consists of an explainable component (Vij) and a random component (eij)

(formula 1). The random component may capture any combination of unobserved attributes, unobserved preference variation, specification error, measurement error and inherent variability within and between individuals (6).

For this analysis we applied dummy coding. For our main effects final model we selected a mixed logit regression model to account for preference heterogeneity with all attributes included as random parameters.

The explainable component (Vij) for this experiment is denoted in formula 2 below, where b1-10 represents the coefficient for the corresponding attribute level. The baseline attribute category for each attribute is omitted from formulae and estimations, as this attribute has by definition a utility of 0 when dummy coding is used.

2. Vij = b1 duration: 2 months + b2 duration: 3months + b3 income loss: 15% + b4 income loss: 25% + b5 Mixed logit models were fit using Stata's mixlogit command which uses simulated maximum likelihood estimators and generates mean utilities for the population and standard deviations of the random coefficients (7) . Mixed logit coefficients (b) can be interpreted as the strength of the relative preference for the particular attribute comparison, with positive coefficients representing positive preferences (desirable) and negative coefficients representing negative preferences (less desirable). Standard deviations represent preference heterogeneity for attribute comparisons, with a 0 standard deviation indicating no heterogeneity.

We further assessed trade-offs by conducting a willingness to trade analysis which is analogous to a traditional willingness to pay analysis (8) . Willingness to pay analyses routinely rely on the assumption of linearity between levels of a continuous attribute (eg. cost, waiting time), given that this assumption of linearity was unlikely to hold beyond the values presented in the experiment, we used nonlinear combinations of estimators 

We fit latent-class conditional logit models through an expectation-maximization algorithm (9) . We fit up to five latent class conditional logit models using maximum likelihood estimation of datasets expanded by sampling weights and selected the model with the smallest model fit criterion (Akaike and Bayesian information criterion), the highest mean probability of group membership and the smallest number of participants with a low probability of group membership in each group. We additionally qualitatively evaluated latent classes to ensure that classes matched heterogeneity demonstrated in main and subgroup analyses. We validated latent class membership using cross-validation techniques (10).

We applied multinomial logit regression to evaluate predictors of latent class membership. Multinomial logistic regression is conducted in the case where a dependent variable is not continuous and has more than two levels -as is the case with 4 latent classes. The model output presents the relative risk ratio which represents the risk ratio between an exposure level and the baseline level for a particular exposure compared to the selected comparison group -in our case the 'risk averse' latent class.

To evaluate marginal probabilities of belonging to the "back to normal" or "risk averse" group we additionally conducted as series of binary logistic regression models with these two categories as the dependent variable and fit an interaction term between gender and other demographic characteristics and generated marginal estimates based on these models, i.e. the probabilities of belonging to each latent class group by gender and demographic characteristic. Multinomial logisitic regression model, with baseline comparison group set to the risk averse.

Getting Started with Conjoint Analysis: Strategies for Product Design and Pricing Research

Conjoint analysis applications in health--a checklist: a report of the ISPOR Good Research Practices for Conjoint Analysis Task Force. Value in health : the journal of the International Society for Pharmacoeconomics and Outcomes Research

Constructing experimental designs for discrete-choice experiments: report of the ISPOR Conjoint Analysis Experimental Design Good Research Practices Task Force. Value in health : the journal of the International Society for Pharmacoeconomics and Outcomes Research

Missouri Census Data Center: Population estimates by age

United States Census Bureau: Missouri

Conducting discrete choice experiments to inform healthcare decision making: a user's guide

Fitting mixed logit models by using maximum simulated likelihood

Willingness to pay for complete symptom relief of gastroesophageal reflux disease

EM Algorithms for nonparametric estimation of mixing distributions

Using Latent Class Analysis to Identify Hidden Clinical Phenotypes