### INTRODUCTION

### DESIGN OF EXPERIMENTS

### Conjoint preference elicitation method

^{3}×4

^{1}) possible combinations. If a respondent is asked to choose one of the two possible profiles or options, total 5,778 (=

_{108}

*C*

_{2}) situations will be possible. In here, we need to design a survey instrument by minimizing the number of choice sets so that respondents feel comfortable to answer without having physical and emotional fatigues and at the same time, this instrument needs to successfully reflect the preferences if 5,778 situations were to be taken from each respondent. In this regard, the orthogonal main effects plan (OMEP) in PE method is most frequently adopted as an optimal solution. The objective of experimental design is to reduce the size of possible choice sets to a manageable number of choice sets for the respondents. Huber and Zwerina2 suggest the 3 properties of OMEP; level balance (equal frequency), orthogonality (independency of the levels of each attribute) and minimal overlap (the probability that an attribute level appears itself in each choice set should be as small as possible). The most common OMEP approach is to use statistical packages such as SPSS or LIMDEP, and this study uses SPSS (it is called 'Orthoplan') to generate an optimal size of choice sets that satisfy three properties of experimental designs. The values in each column indicate the level of corresponding attribute. For each attribute variable, the lowest level is shown as 0 and next level is 1, 2, and 3. Besides, in order to comprise the pairs of choice, we made another option by adding the number generator as explained.

### Attributes and levels

### ESTIMATION STRATEGY (ECONOMETRIC METHODS)

*U*

_{ij}represents the indirect utility function of individual

*i*for good

*j*,

*V*

_{ij}indicates the deterministic component (defined over levels of attributes and observed characteristics) and ε

_{ij}reflects the unobservable factors. With this framework, an individual

*i*will choose

*j*over other alternative of

*k*if

*j*can be shown as below

*V*

_{ij}=

*β'X*. Note that a vector of

*X*is defined over attributes and observable characteristics and β will be empirically estimated.

### Conditional logit model

*ε*-

_{ik}*ε*) will constitute the logistic probability distribution. And the probability that an individual

_{ij}*i*makes choice of

*j*among

*k*alternatives (see equation (3)) can be derived using a closed form of cross-sectional conditional logit equation below where

*V*

_{ij}=

*β'X*.14 Here, a vector of

*X*represents attributes and observed individual characteristics. Additionally, the inclusion of individual characteristics (or socio-economic components) in the estimation leads to a "hybrid" CL models.15

### Mixed logit model

*i*gets from a single choice decision of

*j*can be defined as follows.

*i*which indicates the heterogeneity of preferences across individuals while CL does not. Second, along with the original error term of

*ε*, a new error term of

_{ij}*μ*is added, which captures the characteristics of an individual. So equation (5) is defined as ML model with error-components.20

_{i}'z_{ij}*z*is an observable random vector and

_{ij}*μ*follows

_{i}*N*(0,

*σ*) and varies with individual preferences. The major issue in equation (5) is that the new error term of

_{i}^{2}*μ*+

_{i}'z_{ij}*ε*reflects the dependence of alternatives and heterogeneity across individuals. It means that

_{ij}*μ*explains individual heterogeneity despite the inclusion of

_{i}*ε*that assumes IIA in CL model.

_{ij}*μ*), by integrating the CL probability for all

_{i}*μ*values derived from probability density function of f (

_{i}*μ*|Ω*), we can get the probability in equation (6)

_{i}