This study aims at investigating estimation efficiency of RP/SP models considering both the design of the SP experiment and the assumptions made with respect to error structures. Simulation technique is adopted in this study. In designing the SP experiment, SP attribute values are moved in the direction that changes respondents’ RP behaviour. Specifically, SP attribute values are set X times or 1/X times the RP attribute values. The error structures considered are the following: (a) RP and SP have both a common error component and independent error components (general model); (b) RP and SP have a common error component and SP only has an independent error component (SP-off-RP model); (c) RP and SP have independent error components only (independent model); and (d) RP and SP have a common error component only (double-bound model). Then, the estimation efficiency is evaluated based on the D-error, which is a function of the asymptotic variance covariance matrix. Insights obtained are the following. (1) For the general model, it is difficult to identify an optimal X, but the model offers better estimation efficiency in the neighbourhood of X = 1.0. (2) For the SP-off-RP model and the independent model, the larger the variance in error components between the RP and SP models, the larger the value of X required to minimise D-error.