choice-metrics.com

Hi Community,

I have a question about choice experimental design and choice estimation.

Can we exclude or remove some attributes from the experimental design of a choice model to improve the estimation results? This would mean that the attributes included in the experimental design differ from those in the final model estimation. Or should we keep all attributes from the experimental design?
Please advise.

Thank you in advance.

I believe that there are two fundamental issues with your question.

Issue 1: How can you "improve" estimation results by deleting data? Model fit, such as LL values etc, are only comparable across models using the exact same data. It is not possible to compare model fit if the data is not identical.
Issue 2: If you showed certain attribute levels to respondents in a choice experiment and then omit them in model estimation then you will get biased parameter estimates (unless the attributes that you removed are completely ignored by all respondents, but that is highly unlikely).

So the answer is No. It is not a good idea to remove attributes as this leads to biased parameters and you will also never be able to argue that it improves model fit.

Michiel

Hi Michiel

Thank you for your reply. I have a question about utility functions in relation to design and estimation.
When designing a utility function with specific attributes, should the utility function used for model estimation also include these specific attributes?
Alternatively, is it permissible to use generic attributes for some of them during the estimation? For instance, if the design phase incorporates 19 specific attributes (e.g., b1-b19), can I use a dummy attribute as a generic attribute for all alternatives during the estimation phase?

Design
; alts = BEV, PHEV, ICE
; rows = 36
; block = 6,minsum,noimprov(20 secs)
; bseed = 12345
; con
; eff = (mnl, d, mean)
; bdraws = sobol(5000)
;cond:
if(BEV.PRICEb = 550, BEV.RANGEb = 380),
if(BEV.PRICEb = 820, BEV.RANGEb = 440),
if(BEV.PRICEb = 950, BEV.RANGEb = 500),
if(BEV.PRICEb = 1200, BEV.RANGEb = 500),
if(BEV.PRICEb = 1600, BEV.RANGEb = 600)


; model:


U(BEV) = b1[0.05]+
b2[(n, -0.00001,0.00001 )]*PRICEb[550,820,950,1200,1600] +
b3[(n,0.000001,0.00001)]*RANGEb [380,440, 500, 600] +
b4[(n,-0.00001,0.00001)]*CTIMEb [15, 30, 60] +
b5.dummy[(n,0.00001,0.00001)|(n,0.00001,0.00001)]*CLOCATION[3,2,1] +
b6.dummy[(n,0.00001,0.00001 )|(n,0.00001,0.00001)]*CDENSITY[3,2,1] +
b7.dummy[(n,0.00001,0.00001)]*CHOME[2,1] +
b8.dummy[(n,0.00001,0.00001)|(n,0.00001,0.00001)|(n,0.00001,0.00001)]*POLICY[4,3,2,1] /

U(PHEV) = b9 [0.05]+
b10[(n,-0.00001,0.00001)]*PRICEp[880,980,1200, 1400,1600] +
b11[(n,0.000001,0.00001)]*RANGEp [440, 540, 680, 760] +
b12[(n,-0.00001,0.00001)]*CTIMEp [5, 10, 15] +
b13.dummy[(n,0.00001,0.00001)|(n,0.00001,0.00001)]*CLOCATION[3,2,1] +
b14.dummy[(n,0.00001,0.00001 )|(n,0.00001,0.00001)]*CDENSITY[3,2,1] +
b15.dummy[(n,0.00001,0.00001)]*CHOME[2,1] +
b16.dummy[(n,0.00001,0.00001)|(n,0.00001,0.00001)|(n,0.00001,0.00001)]*POLICY[4,3,2,1] /

U(ICE) = b17[(n,-0.00001,0.00001)]*PRICEi[530,630,870,990,1200] +
b18[(n,0.00001,0.00001)]*RANGEi [640,760, 840, 920] +
b19[(n,-0.00001,0.00001)]*CTIMEi [3, 5, 10]

$

When you say a utility function with "specific attributes" and when you refer to "generic attributes", do you mean "alternative-specific coefficients/parameters" and "generic coefficients/parameters"? I think that you may be mixing up terminology. An attribute is a characteristic of an alternative, such as CHOME, whereas a coefficient is the unknown behavioural parameter that you aim to estimate, such as b7.

If you indeed refer to coefficients/parameters instead of attributes, then the answer is YES. You can use generic coefficients instead of alternative-specific coefficients for your attributes during model estimation. For example, you could conduct a statistical test, H0: b7-b15 = 0 (there are two dummy coefficients you there will be two tests). If you are unable to reject this null hypothesis then you can use a generic coefficient for CHOME across BEV and PHEV. If you reject the null hypothesis then it is typically not a good idea to make the coefficient generic across alternatives. You could also consider a generic coefficient for PRICEb and PRICEp, even if they have different attribute levels.

Michiel