choice-metrics.com

Hi all,

I've been building a labeled choice experiment. In my CE, respondents are provided with six labeled choice alternatives, which are descriptions of ecological main heating systems. These alternatives are district heat (reference category in the analysis, only alternative specific constant calculated), solid wood, wood pellet, electric storage heating, ground heat and the exhaust air heat pump. The alternatives to choose from are described by the following five attributes: supplementary heating systems (sup), investment costs (icost), operating costs (ocost), comfort of use (comf) and environmental friendliness (enfrend). All these attributes, except supplementary heating system, have label -specific attribute levels. For instance, the investment cost for ground heat pumps has four levels 13000€, 16000€, 19000€ and 22000€, whereas the investment cost for solid wood fired has different four levels 4500€, 7000€, 9500€ and 12000€.

The syntax I’ve been using is following:
Design
? Note that orthogonal designs generated around 70 choice tasks at minimum --> This is too much for my sample size ?
? prior values have to be close to zero to be able to find any efficient design whatsoever ?
? District heat works as a reference category and it is taken into account when planning the overall design ?
? alternatives = ground, exair, wood, pellet, elect?
? Note that the prior values are correct only in terms of the sign ?
; alts = ground, exair, wood, pellet, elect
; rows = 32
; eff = (mnl,d)
; block = 4
; model:
U(ground)= sup.effects[0.02|0.02|0.02]*sup[0,1,2,3] +icost[-0.02]*gricost[13000,16000,19000,22000] + ocost[-0.01]*grocost[500,650,800,950] + comf[0.02]*bettercomf[1,2] + enfrend[0.01]*betterenfrend[1,2]/
U(exair)= sup.effects *sup + icost[-0.02]*exicost[7000,9000,11000,13000] + ocost[-0.01]*exocost[800,1000,1200,1400] + comf*bettercomf + enfrend*worseenfrend[0,1]/
U(wood)= sup.effects *sup + icost[-0.02]*woicost[4500,7000,9500,12000] + ocost[-0.01]*woocost[600,850,1100,1350] + comf*worsecomf[0,1] + enfrend*betterenfrend/
U(pellet)= sup.effects *sup + icost[-0.02]*peicost[8000,11000,13000,16000] + ocost[-0.01]*peocost[750,950,1150,1350] + comf*worsecomf + enfrend*betterenfrend/
U(elect)= sup.effects *sup + icost[-0.02]*elicost[6000,8500,11000,13500] + ocost[-0.01]*elocost[1500,1700,1900,2100] + comf*bettercomf + enfrend*worseenfrend
$

I have few questions about my design and application.
Firstly, how should I treat these alternative specific attributes (icost, ocost, comf and enfrend)? At this stage, I’ve used same coefficients to link the alternatives with varying attribute levels together. Is there some better way to do this? Should I construct the whole design in a way that I “reject” those attribute levels which aren’t relevant for the alternative?
Secondly, why the prior values have to be close to zero to find any kind of efficient design?

Thanks already in advance for your comments.

Best Regards,
Enni

I am not really sure I understand your questions.

1. You can use generic or alternative specific coefficients depending on whether you think what is the model you will estimate in the end. You can easily use icost_ground and icost_exair with different priors, you just get more parameters to estimate. But you can keep them the same, even if the attribute levels are different. For each alternative, you only put the levels in that are relevant for that alternative, so there is nothing to reject.

2. Since your attribute levels are very large (16000), and your utility functions look like beta * attribute, beta * 16000 will be a very large value if your beta is not very small. So this attribute would completely dominate your choice task and hence you will not be able to estimate the model. In all cases, the priors need to make sense, and it does not seem that you have chosen priors that either come from the literature or that make sense behaviourally, you seem to have randomly put small numbers there. For the attribute bettercomf with levels 1 and 2, you will likely have higher priors (not 0.02 but rather something like 0.2), while for gricost with levels up to 22000 you will likely have lower priors (not 0.02 but rather something like 0.0001). These priors preferably come from estimation using data from a pilot study. You should choose your priors very carefully, as they have a huge impact on the efficiency of your design. In case you have no knowledge on them, you can set them to zero.

Thanks a lot for your answer! It was really helpful.