Random Utility Maximization Model

The Random Utility Maximization framework (RUM) assumes individuals are utility maximizers, where utility is derived from the attributes of the product or good being consumed (48). Given that the researcher only observes the choice of the consumer, it is assumed that the individual's utility function is represented as Unit=β′Xnit+εnit, where X is a vector of attributes observed for individual n (1, …, N) when choosing alternative i (1, …, I) in choice set t (1, …, T), β is a vector of parameters to be estimated, and ε is the unobserved part of utility. Given that ε is unobserved, we treat ε as a random variable. Following Train (49), we assume ε is distributed Extreme Value Type I. Train (49) explains that this distributional assumption does not differ substantially from the normal distribution, but does allow for more aberrant behavior given it has fatter tails. In addition, this distributional assumption allows for a closed-from solution for the choice probabilities of interest in this study. The probability of an individual choosing alternative i over alternative j in choice set t is given by the Random Utility Multinomial Logit (RUM-MNL) function:

where Vnit=β′Xnit.

The RUM-MNL assumes that all individuals have similar views on food safety recall attributes. It is likely that these views vary across individuals and that groups of individuals have similar tastes and preferences. We allow for this type of heterogeneity using a latent class model in which attribute estimates are assumed to be similar within groups/classes but different across groups/classes. The approach and number of classes to include was determined using a combination of the Akaike information criterion [AIC; (50)], adjusted Bayesian information criterion [BIC; (51)], and relevant class sizes consisting of at least 20% of the individuals (52, 53). Thus, we modify Equation 1 to allow estimation of classes, C, given as:

where P_nc is the probability of individual n being in class c, V_nit is the same as described above expect now the parameters are class specific (i.e., β_c). P_nc is assumed to be a function of individual-specific characteristics that describe the characteristics of the group. That is: Pnc=exp(γc′Zn)∑r=1Rexp(γr′Zn), where γ_c is a vector of class parameters to be estimated and Z_n is a vector of individual specific covariates. We assume class membership is determined by trust in informational sources, buying behavior, and cooking behavior (6–8).