Appendix 1: Estimation of the two-stage model
We discuss the estimation method in this appendix. We assume that the utilities in the consideration set stage and the choice stage are interrelated and cast the two utility functions into a system of equations as follows:
$$ \left( {\matrix{ {Z_i^{ * }} \hfill \\ {Y_i^{ * }} \hfill \\ }<!end array> } \right) = \left( {\matrix{ {X_i^C} \hfill & 0 \hfill \\ 0 \hfill &{X_i^F} \hfill \\ }<!end array> } \right)\left( {\matrix{ \beta \\ \delta \\ }<!end array> } \right) + \left( {\matrix{ {{\eta_i}} \hfill \\ {{\varepsilon_i}} \hfill \\ }<!end array> } \right), $$
(1)
where \( Z_i^{*} \) is the vector of utilities from the consideration set stage and \( Y_i^{*} \) is the vector of utilities from the choice stage. We assume that \( \left( {\matrix{ {{\eta_i}} \\ {{\varepsilon_i}} \\ }<!end array> } \right)\sim N\left( {\left( {\matrix{ 0 \\ 0 \\ }<!end array> } \right),\sum { = \left( {\matrix{ {{\Sigma_{{zz}}}} \hfill &{{\Sigma_{{zy}}}} \hfill \\ {\Sigma_{{zy}}^{\prime }} \hfill &{{\Sigma_{{yy}}}} \hfill \\ }<!end array> } \right)} } \right) \), where ∑
zz
and ∑
yy
are variance–covariance matrixes of η
i
and ε
i
, respectively, and ∑
zy
is the covariance matrix between η
i
and ε
i
.
Thus, our two-stage model is given by Eq. 1. It allows different parameters for consideration utility \( Z_i^{*} \) and choice utility \( Y_i^{*} \), which is a flexible representation (Gilbride and Allenby 2004). And, by considering ∑
zy
, we can reflect the relationships between consideration set utility and choice utility in estimation, which is relatively unexamined in the literature because it is empirically difficult to model correlation between the two stages (Van Nierop et al. 2010).
Equation 1 is in the form of a SUR model, and the estimation method for parameters β, δ, and ∑ is discussed in several places (e.g., Koop 2003). The difference from a standard SUR model is that we need to draw \( Z_i^{ * } \) and \( Y_i^{ * } \) by data augmentation using the consideration set and the choice. We draw \( Z_i^{ * } \) using the data augmentation method of a multivariate probit model because consumers decide for each of the J alternatives whether it should be included or not. However, we draw \( Y_i^{ * } \) using the data augmentation method of a multinomial probit model because consumers choose a specific product from the consideration set. The full Bayesian MCMC algorithms including the data augmentation procedure are in “Appendix 2,” which is available from the authors.
The probability that consumer i chooses product j from consideration set C
i
in the choice stage can be written as \( P\left( {{y_i} = j} \right) = P\left( {{y_i} = j\left| {{C_i}} \right.} \right) \times P\left( {{C_i}} \right) \). Here, P(C
i
) is the probability of observed consideration set C
i
, and it is calculated as follows:
$$ P\left( {{C_i}} \right) = P\left( {{c_{{i1}}},{c_{{i2}}}, \cdots, {c_{{iJ}}}} \right) = P\left( {{d_{{i1}}}z_{{i1}}^{ * } \geqslant 0,\quad {d_{{i2}}}z_{{i2}}^{ * } \geqslant 0,\quad \cdots, \quad {d_{{iJ}}}z_{{iJ}}^{ * } \geqslant 0} \right), $$
where c
ij
= 1 if consumer i includes product j in the consideration set and c
ij
= 0 if not. And d
ij
= 1 if consumer i includes product j in the consideration set and d
ij
= −1 if not. For example, if consumer i includes products 1 and 2 out of three products {1, 2, 3}, the probability of consideration C
i
is calculated as \( P\left( {{C_i}} \right) = P\left( {{c_{{i1}}} = 1,{c_{{i2}}} = 1,{c_{{i3}}} = 0} \right) = P\left( {z_{{i1}}^{ * } \geqslant 0,\quad z_{{i2}}^{ * } \geqslant 0,\quad z_{{i3}}^{ * } \leqslant 0} \right) \).
Furthermore, P(y
i
= j | C
i
) is the probability that consumer i finally chooses product j given the consideration set C
i
, and it is calculated as follows:
$$ P\left( {{y_i} = j\left| {{C_i}} \right.} \right) = P\left( {y_{{ij}}^{ * } \geqslant y_{{ik}}^{ * }} \right)\forall \;k \ne j\left| {j,k \in {C_i}} \right., $$
where y
i
is consumer i’s choice among J products. We calculate this probability using the GHK estimator.
Equation 1 involves a multivariate probit model for consideration set and a multinomial probit model for the choice. As these are discrete choice models, we cannot identify all the parameters in Eq. 1. Following Edwards and Allenby (2003), we navigate in the unidentified parameter space in estimation but report parameters which are divided by the corresponding variances.