| Abstract: | A feasible multinomial estimation procedure is derived, which does not require parameterization of the elements in the covariance matrix. The estimation is carried out using a simulated expectation‐maximization algorithm, where the covariance structure is evaluated based on a set of score functions, while the structural coefficients are derived using standard multinomial probit (MNP) conditional on the given covariance structure. This methodology is demonstrated using a Monte Carlo simulation on both rank‐ordered and non‐ranked data, and on a real data set involving the choice of local residential telephone service. For limited finite samples, the procedure is shown to be superior to conventional MNP since it is faster, involves fewer parameters, and generates estimates with smaller variances. |
