The Location Method

This section will detail the experimental design and simulation process employed in the first phase of the survey which used the location method of attribute representation. This method of attribute representation was formed with attribute levels based in the context of apartment location in the block.

The survey conducted by Arsenio (2002) revealed that the average number of floors in a block in the area were six. Based on this information, the levels of the attributes using the location method were set. Thus under the location method, view, noise and sunlight were categorised in to four levels based on their location in the block. A slight variation from the method employed by Arsenio (2002) was carried out for this attribute representation method. On the basis of the noise levels obtained for the different floors in that study, the locations on the sixth floor (front and back façade of the block in relation to the main traffic road, thus 6F – apartment on the sixth floor facing the main traffic road and 6T – apartment on the sixth floor facing in the opposite direction of the main traffic road) and similarly those on the third floor (3F and 3T) were selected as levels for the choice experiment in the current study.

As pivotal design and computer aided interview were not employed in the current survey due to resource limitations, the levels for the third and the sixth floor of the block were fixed in the experimental design as these floors showed marked variation in noise levels as given in Arsenio (2002). Moreover, an average of six floors in a block in the area as stated by Arsenio (2002) implied that sufficient blocks would be available for the survey if the sixth floor was chosen as an upper floor level. While the absence of considering the respondent’s current apartment location in the experimental design can be regarded as a design limitation, this method was employed considering the resource availability for the survey.

Moreover, care was also taken that the residential blocks selected for the survey contained at least six floors in order to keep the choice scenarios as realistic as possible.

A fractional factorial orthogonal design based on attribute differences was developed using the experimental plan given by Kocur et al. (1982) and explained in section 5.3.

The four apartment locations used in the experimental design can be summarised as follows:

Table 5.1 Apartment location description

Levels Description

6F Apartment on 6^thfloor on façade facing the main traffic road 6T Apartment on 6^thfloor on façade opposite the main traffic road 3F Apartment on 3^rdfloor on façade facing the main traffic road 3T Apartment on 3^rdfloor on façade opposite the main traffic road

From the information gathered from the blocks as well as from the estate agencies in the area, it was gathered that the average charge value in the area was between 40 – 60 Euros per month. Based on this figure and in order to have the signs of the charge difference evenly distributed across the choice scenarios, both positive and negative signs of charge difference were employed in the experimental design. The levels of all the attributes in the experimental design can be given as follows:

Table 5.2 Experimental design levels for view, noise and sunlight

-The fractional factorial orthogonal main effects design code for four attributes and four levels, obtained from Kocur et al. (1982) gave 16 choice scenarios in total which were split into two different sets of choice scenarios based on the choice elicitation method used and as summarised in Section 5.3 and in Figure 5.2.

Simulation was carried out with different levels of the charge difference as well as different coefficient values (including coefficient values obtained from Arsenio,

2002). As the perception ratings for view, noise and sunlight levels were not already known, random numbers were generated for these between ranges (20-80) as well as (0-100) for the simulation process. Though different levels of charge differences were experimented during the simulation process, levels that made practical sense in terms of the absolute values as well as the charge levels prevalent in the area were chosen. Simulation was carried out with this level to test the goodness of model fit and the statistical significance of the coefficient values.

Different values of standard deviation (s.d.) during the simulation process were also experimented. Table 5.3 gives the results obtained from the simulation with different levels of attribute values and attribute coefficient with the standard deviation of 1.28 and 2.56.

The variations between the different types of models run during the simulation were based on different attribute and coefficient values. Each of the models given in the table has two different input values (20-80) and (0-100). Thus model 1 has one set of input coefficient and two different sets of input values while model 2 again has the same sets of input values but different input coefficients from model 1. Effect of different standard deviation in simulation has also been given here for some input values and coefficient. Thus, models 5, 7 and 10 provide the effect of standard deviation 2.56 compared to models 4, 1 and 8 respectively which apply the standard deviation of 1.28.

Table 5.3 Simulation results for location method with different input and coefficient values

INPUT COEFFICIENT ESTIMATED COEFFICIENT

V N C S s.d V (t-ratio) N C S Adj. ρ²

Model 1 Input (20-80) .024 .031 -.015 .017 1.28 .026 (10.7) .028 (11.6) -.014 (-7.9) .012 (5.4) .153 Input (0-100) .024 .031 -.015 .017 1.28 .025 (13.7) .032 (16.9) -.016 (-7.7) .017 (10.1) .32 Model 2 Input (20-80) .012 .015 -.007 -.008 1.28 .014 (6.6) .013 (6.3) -.007 (-4.7) .009 (4.4) .053

Input (0-100) .012 .015 -.007 -.008 1.28 .012 (8.7) .017 (11.9) -.008 (-4.9) -.008 (-5.8) .126 Model 3 Input (20-80) .012 .031 -.015 .017 1.28 .015 (6.6) .032 (13.1) -.013 (-7.7) .014 (6.4) .142 Input (0-100) .012 .031 -.015 .017 1.28 .014 (8.8) .031 (16.4) -.014 (-7.2) .019 (11.9) .267 Model 4 Input (20-80) .048 .24 -.2 -.03 1.28 .039 (7.3) .19 (14.9) -.16 (-15.3) -.028 (-5.7) .78

Input (0-100) .048 .24 -.2 -.03 1.28 .038 (9.5) .18 (14.3) -.15 (-13.9) -.019 (5.8) .805 Model 5 Input (20-80) .048 .24 -.2 -.03 2.56 .021 (5.6) .12 (17.2) -.10 (-18.5) -.009 (-2.7) .63

Input (0-100) .048 .24 -.2 -.03 2.56 .026 (8.7) .12 (16.3) -.10 (-15.3) -.014 (-5.8) .73 Model 6 Input (20-80) .024 .12 -.1 .03 1.28 .024 (6.6) .12 (16.6) -.098 (-18) .03 (8.0) .62 Input (0-100) .024 .12 -.1 .03 1.28 .024 (8.2) .12 (16.6) -.09 (-15.9) .027 (9.2) .715 Model 7 Input (20-80) .024 .031 -.015 .017 2.56 .009 (4.0) .017 (7.66) -.0096 (-6) .007 (3.4) .054 Input (0-100) .024 .031 -.015 .017 2.56 .012 (8.8) .018 (12.7) -.004 (-2.4) .008 (5.8) .127 Model 8 Input (20-80) .048 .062 -.03 .034 1.28 .049 (14.8) .063 (16.8) -.028 (-12) .034 (11.3) .382 Input (0-100) .048 .062 -.03 .034 1.28 .046 (16.6) .061 (18.1) -.028 (-10) .034 (13.6) .557 Model 9 Input (20-80) .024 .031 -.015 .008 1.28 .026 (10.9) .030 (11.9) -.012 (-6.9) .012 (5.3) .163 Input (0-100) .024 .031 -.015 .008 1.28 .022 (13.2) .032 (16.8) -.017 (-8.5) .0085 (5.7) .276 Model 10 Input (20-80) .048 .062 -.03 .034 2.56 .025 (10.6) .028 (11.7) -.014 (-7.6) .0156 (6.8) .158 Input (0-100) .048 .062 -.03 .034 2.56 .023 (13.0) .031 (16.0) -.015 (-7.6) .021 (12.4) .320

Based on the results obtained from the simulation, it can be seen that the estimated coefficient values are close to the input values and are also statistically significant in case of standard deviation of 1.28 (which is the assumption of the MNL model).

In case of standard deviation of 2.56, the effect of using that value can be observed in the coefficient estimates which show an almost halving of the input coefficient values used. Using the input values, it is seen that most of the models give an acceptable fit in terms of rho-square values (an acceptable fit is considered to be between 0.1 – 0.2). As the values of the attributes view, noise and sunlight are unknown due to its dependence on respondent’s perception ratings, the main aim of the simulation was to determine whether the levels of charge difference selected gave reasonable results. The results obtained indicate that this is the case for almost all the specifications experimented. Based on the simulation exercise, it can thus be concluded that the design developed is fit for purpose.