Dear Statalist,
I would much appreciate your help!
I am analysing an unlabelled choice experiment on employee choice with an opt out alternative.
So each respondands answeres six choice sets by choosing one of 3 alternatives ( A B & optout). I have three kategorical and two metric variables (e.g. fulltime position / 80% position and commute in minutes). Whereby the opt out alternative always is assigned the value 0 (to be clear: 0 (optout) 1(fulltime) 2 (80% position) ). I have 1,499 respondands, 9000 choice sets and 26 982 valid oberservations.
Here is an example of three choice sets from id = 100
My main research objective is to analyse valuation heterogeneity.
The basic model and it’s interpretation work fine, but I am not sure how to proceed when analysing group differences.
Basically I see two strategies, but as far as I comprehend the literature, both are not statistically valid or at least frowned upon.
I could compute asclogit/clogit models seperatly for each subgroup (e.g. gender) and then compare estimation coefficients. In the case of non-linear modells with binary dependant variable this is not possible without further analyses, as the computation of effect coefficents depend on unobserved heterogeneity. So usually, in logit estimations, one would compare average marginal effects between groups. So I computed:
Can someone explain what I am missing?
Which transformation of the c-logit effects do I have to perform to be able to interpret differences between group-specific models?
The other altenative would be to include interaction effects between personal characteristics and attributes (eg. Workhrs_female = 1 if - female == 1 & workshrs == 2 (fulltime), = 0 in all opt out alternatives, all male respondands and all female alternatives sets with workhrs = 1 part time). I have read frequently, that such interaction effects are often misinterpretated in simple logit modells.
I am therefore unsure if it would be valid computation to include such interactioneffects. Would the interpretation be as simple as this (Let`s say OR of workhrs : 0.75 OR workhrs_female: 0.80)
A fulltime alternative has a decreased odds of choosing that employment by the factor of 0.75 for men, compared to an 80% employment and holding all other alternatives constant. For Women the decrease of odds is by the factor 0.6 (0.75*0.80).
So both genders prefere part time employment, but for women this attribute is more relevant in the decision making process?
I have thought about using nlogit but I do not really see an advantage of calculating the opt out alternative as a separate branch.[INDENT=2] type N option N k
---------------------------------------
alt 18000 --- Option A 9000 3412
Option B 9000 3538
none9000 --- none of th~e 9000 2050
---------------------------------------
[/INDENT]Thank you very much in advance. All help and comments are highly appreciated.
Sara
I am using Stata/MP 15.1.
I would much appreciate your help!
I am analysing an unlabelled choice experiment on employee choice with an opt out alternative.
So each respondands answeres six choice sets by choosing one of 3 alternatives ( A B & optout). I have three kategorical and two metric variables (e.g. fulltime position / 80% position and commute in minutes). Whereby the opt out alternative always is assigned the value 0 (to be clear: 0 (optout) 1(fulltime) 2 (80% position) ). I have 1,499 respondands, 9000 choice sets and 26 982 valid oberservations.
Here is an example of three choice sets from id = 100
id str choice income workhrsAs I have unlabeled alternatives the asclogit and the clogit command give me the same estimation results.
1. 100 101 0 400 1 fulltime
2. 100 101 1 400 2 80%
3. 100 101 0 0 0
4. 100 102 0 400 1 fulltime
5. 100 102 1 500 1 fulltime
6. 100 102 0 0 0
7. 100 103 1 500 2 80%
8. 100 103 0 700 1 fulltime
9. 100 103 0 0 0
My main research objective is to analyse valuation heterogeneity.
The basic model and it’s interpretation work fine, but I am not sure how to proceed when analysing group differences.
Basically I see two strategies, but as far as I comprehend the literature, both are not statistically valid or at least frowned upon.
I could compute asclogit/clogit models seperatly for each subgroup (e.g. gender) and then compare estimation coefficients. In the case of non-linear modells with binary dependant variable this is not possible without further analyses, as the computation of effect coefficents depend on unobserved heterogeneity. So usually, in logit estimations, one would compare average marginal effects between groups. So I computed:
clogit decision alt2 alt3 inc100 retain commute10 workhrs desorientation , group (str) vce(cluster newid)The clogit coef. and the margins dy/dx are identical. This confuses me, because I thought that the clogit coefs are interpreted indentically to the logit command as change of odds, whereas margins estimated change in probability.
margins, dydx(*) predict(xb)
Can someone explain what I am missing?
Which transformation of the c-logit effects do I have to perform to be able to interpret differences between group-specific models?
The other altenative would be to include interaction effects between personal characteristics and attributes (eg. Workhrs_female = 1 if - female == 1 & workshrs == 2 (fulltime), = 0 in all opt out alternatives, all male respondands and all female alternatives sets with workhrs = 1 part time). I have read frequently, that such interaction effects are often misinterpretated in simple logit modells.
I am therefore unsure if it would be valid computation to include such interactioneffects. Would the interpretation be as simple as this (Let`s say OR of workhrs : 0.75 OR workhrs_female: 0.80)
A fulltime alternative has a decreased odds of choosing that employment by the factor of 0.75 for men, compared to an 80% employment and holding all other alternatives constant. For Women the decrease of odds is by the factor 0.6 (0.75*0.80).
So both genders prefere part time employment, but for women this attribute is more relevant in the decision making process?
I have thought about using nlogit but I do not really see an advantage of calculating the opt out alternative as a separate branch.[INDENT=2] type N option N k
---------------------------------------
alt 18000 --- Option A 9000 3412
Option B 9000 3538
none9000 --- none of th~e 9000 2050
---------------------------------------
[/INDENT]Thank you very much in advance. All help and comments are highly appreciated.
Sara
I am using Stata/MP 15.1.
