Monetary savings coefficient with unexpected negative sign

[tomschuette]

Dear all,

I have conducted a stated choice experiment using a Choice Design from Ngene. It all worked very well, with one exception. One of the estmated parameters has the "wrong" sign, and I have no explanation for it. I am now concerned that it might be due to a design issue and not due to actual preferences from the respondents.

The syntax used was the following:

Code: Select all

Design
;alts = opt1*, opt2*, opt3*, opt_out
;rows = 180
;block = 30
;eff = (mnl, d)
;alg = mfederov

;reject:

opt1.pv = 1 and opt1.costs > 60000,
opt2.pv = 1 and opt2.costs > 60000,
opt3.pv = 1 and opt3.costs > 60000,

opt1.pv > 1 and opt1.costs < 50000,
opt2.pv > 1 and opt2.costs < 50000,
opt3.pv > 1 and opt3.costs < 50000,

opt1.pv = 1 and opt1.savings > 20,
opt2.pv = 1 and opt2.savings > 20,
opt3.pv = 1 and opt3.savings > 20,

opt1.pv = 2 and opt1.savings < 20,
opt2.pv = 2 and opt2.savings < 20,
opt3.pv = 2 and opt3.savings < 20,
opt1.pv = 2 and opt1.savings > 35,
opt2.pv = 2 and opt2.savings > 35,
opt3.pv = 2 and opt3.savings > 35,

opt1.pv = 3 and opt1.savings < 35, 
opt2.pv = 3 and opt2.savings < 35, 
opt3.pv = 3 and opt3.savings < 35, 
opt1.pv = 3 and opt1.savings > 45,
opt2.pv = 3 and opt2.savings > 45,
opt3.pv = 3 and opt3.savings > 45,

opt1.pv = 4 and opt1.savings < 45,
opt2.pv = 4 and opt2.savings < 45,
opt3.pv = 4 and opt3.savings < 45


;model:

U(opt1) = b1.dummy[0.0001]                       *sustainablematerial[2,1]
        + b2.dummy[0.0000005|0.000001]           *energyefficiency[2,3,1]
        + b3.dummy[0.000025|0.00005|0.000075]    *pv[2,3,4,1]
        + b4.dummy[-0.00005|-0.0001]             *manufacturer[2,3,1]
        + b5[0.0001]                             *savings[5, 10, 20, 30, 35, 40, 45, 50, 55]
        + b6[-0.0001]                            *costs[30000, 40000, 50000, 60000, 70000, 80000, 90000]

/
U(opt2) = b1*sustainablematerial + b2*energyefficiency + b3*pv + b4*manufacturer + b5*savings + b6*costs
/
U(opt3) = b1*sustainablematerial + b2*energyefficiency + b3*pv + b4*manufacturer + b5*savings + b6*costs

$

The attribute of interest is the "savings" attribute, which through the "reject" statements (with level overlap) is set to be, on average, higher with higher levels of the attribute "pv" to make the alternatives more realistic. The estimated parameter in a mixed logit model is now statistically significant and negative, leading to the interpretation, that, ceteris paribus, respondents prefer lower levels of monetary savings on their energy bill, which is highly unrealistic. All other parameters have the expected signs.

Are there any reasons to believe that this is caused by the design created through the syntax above?

Thanks in advance for any insights and comments.

[Michiel Bliemer]

It could be that people have strong preferences towards PV and since you imposed come constraints between PV and SAVINGS, the resulting correlation may affect the design of SAVINGS. But I think that this would only occur if PV is a much more important attribute than SAVINGS.

Another issue I notice in your design is that you did not include a constant for the opt-out and that you set the prior for COSTS relative large, namely -0.0001 relative to attribute levels that are in their ten thousands. In other words, this creates a very dominant attribute COSTS and if you look at the choice probabilities of each choice task, the opt-out is vastly preferred (99%) in all cases since the COSTS attribute imposes a very large negative utility on opt1-opt3 while the opt-out has utility 0. This results in a quite inefficient design and I am not sure what the impact of this will be.

Michiel