Priors and ASC in labelled design

[Rajwane]

Hello,

I am a new user of Ngene and this is my first DCE. I am working on a labeled DCE to estimate WTPs for repairing household appliances.

I have 4 alternatives : DIY repair, Pro repair, Replace with new, and opt-out option.

I have already done the first design, and the pilot study. So I want to add the priors I got from my mnl estimation in stata. However, I have a few questions :

- Are the priors the same for all alternatives ? means if the beta coefficient for attribute cost is 0.3, do I use the same prior for all of my three alternatives ?

- Concerning the ASC, I know that it is better to add them.. I did not in the pilot study design however. If I understood well from my readings on this forum, I can add ASC to only three of the four alternatives I have. So the prior for the ASC is what in this case ? In the database collected from the pilot study I have added dummy coded variables for each alternative... do I use this coefficient as a prior ?

- Do I need to take other things into consideration ?

-

Code: Select all

design
;alts = DIY*, PRO*, Replace*, neither
;rows = 24
;block = 3
; eff = (mnl, d)
;cond:
if(DIY.cost1=10, PRO.cost2=[50,85]),
if(DIY.cost1=40, PRO.cost2=[85,120,200]),
if(DIY.cost1=70, PRO.cost2=[120,200]),
if(DIY.cost1=150, PRO.cost2=[200]),
if(PRO.cost2=50, PRO.bonus=[15]),
if(PRO.cost2=120, PRO.bonus=[15,30])

;model :
U(DIY) = DIY[0]
       + b1.dummy[0.2]*tutorial [1,0] 
       + b2[-0.00149]*cost1 [10,40,70,150]
       + b4[-0.0344]*CO2 [10,20]
       + b5[-0.03]*duration1 [3,7,15]/

U(PRO)= PRO[0]
      + b2[-0.00149]*cost2 [50,85,120,200]
      + b3[0.00001]*bonus [15,30]
      + b4[-0.0344]*CO2 [10,20]
      + b5[-0.03]*duration2 [1,3,7,15]
      + b6[0.0592]*warranty1 [3,6,12]/

U(Replace) = b2[-0.00149]*cost3 [300,400,500]
            + b4[-0.0344]*CO2 [10,20]
            + b5[-0.03]*duration3 [3,7]
            + b6[0.0592]*warranty2 [24]/

U(neither) = neither[0]

$


[Michiel Bliemer]

With labelled alternatives, parameters can be generic across all alternatives, or can be alternative-specific. For example, if you expect that sensitivity to cost is different for DIY and PRO then you would use alternative-specific parameters, e.g. b2_diy and b2_pro, and they would have different priors that come out of your pilot study. However, if you expect that cost sensitivity is generic, then you would use b2 for both alternatives and you use the same prior. In most cases, cost has a generic parameter (economists argue that cost = cost, no matter what you spent it on), but in some cases the parameters could be alternative-specific. For example in transport, the parameter for travel time in a bus is usually different from the parameter for travel time in a car because travel time is experienced very differently (e.g. in the car one may need to drive and pay attention to the road, while in a bus one could play games on a phone).

Regarding your second question, you would need to estimate the constants in your pilot data. You can add these constants to the utility function during model estimation, or you could create a dummy-coded variable that describes the utility attached to each of the labels of the alternatives. You seem to have done the latter (with Replace being the base level since its constant is normalised to zero), so then you would indeed use the priors for the parameters of the dummy coded label variable as priors for the constants.

Two other comments on your script:
- Remove the asterisk (*) from your alternatives since your alternatives are not unlabelled
- Add ;con to your script if you want to also optimise the design for estimating the ASCs

Michiel

[Rajwane]

Thank you so much for your reply Dr. Bliemer.

I estimated the following equation on stata : Choice = dummyAlt1 + dummyAlt2 + list of attributes

The dummies are simply dummy variables that I created in the database.

The coefficients I estimated using stata for these dummy variables are very large compared to the other coefficients (9.969*** for the first alternative and 9.228*** for the second alternative). The D-error jumped to 21.5 when creating the design, so i replaced them by 0.969 and 0.9228 respectively. So the D-error is now 0.016. Is that even possible to decrease the value of the coefficient ? or does this mean that there is an error somewhere ?

Here are the beta coefficients estimated with Stata :

Autorepairing (Alternative 1) 9.969***(2.286)
Professionnal (Alternative 2) 9.228***(2.357)
Manuels or tutorials at disposal(b1) -0.274(0.218)
Cost (b2) -0.00149***(0.000524)
Repair bonus (b3) 0.0118 (0.00933)
CO2 emissions (b4) 0.0344***(0.00960)
Post-repair warranty (b5) 0.0592*(0.0324)
Repair duration (b6) 0.0300*(0.0155)
Constant -6.402***(1.358)

Thank you so much for your help

[Michiel Bliemer]

Having two constants with such large size compared to Replace with a constant of 0 does not make much sense, unless people in the data never selected Replace (perhaps have a quick look). If people do select Replace regularly then there seems to be something wrong with your model estimation or coding of your data. I cannot really help there as I have no expertise with model estimation using Stata. You should not have to manually adjust your coefficients, including your constants, that much.

Michiel