Dear all,
I am currently working on a dataset of a forced choice based conjoint experiment and are a little bit troubling on how to analyse the data.
Currently, I am using a mixed-effects model, using the following code:
melogit dependent_var charA1 charA2 charB1 charB2 charC1 charC2 || respondent_ID:
I am currently not sure, if this is correct. The experiment was conducted as follows:
The respondents were presented two objects, that differed in 3 characteristics, and they had to decide, which one they would choose. If they chose object A, this would mark the dependent variable as "1" (made a decision for object A), and simultaneously, the other object that was presented, was marked as dependent_var "0" (did not chose object B). Respondents had to do this 13 times, leading to 26 decisions in total (13 "would chose" decisions, where the dependent variable was then a "1", and 13 "would not chose" decisions, where the dependent variable was then a "0").
The objects differed in three characteristics. These characteristics were always allocated randomly and are coded as dummy variables. This leads to data, that somehow looked as follows:
dependent_var = 1 ; charA1 = 0; charA2 = 1; charB1 = 1; charB2 = 0; charC1 = 1; charC2 = 0; respondent_ID = 45
dependent_var = 0 ; charA1 = 1; charA2 = 0; charB1 = 0; charB2 = 1; charC1 = 0; charC2 = 1; respondent_ID = 45
....(there were 24 more observations for respondent 45, but I cut this here for the sake of brevity).
(This is a little bit reduced, in fact, there were seven characteristics, where some had four different values, and some had three different values).
So these 26 decisions are actually "nested" inside the respondents ID. With logistic regression, I am now trying to find out, which characteristics are the most important ones for the respondents when they want to make a decision. Because of the nested data (I had 300 respondents, each of them made 26 decisions), I want to use a multi-level regression.
Can anyone tell me, if the approach that I described in the beginning, is meaningful? I am also considering to append vce(cluster respondent_ID) to avoid biases according to the small first-level data (i.e., only 26 per respondent), which unfortunately decreases my z-values and makes some coefficients insignificant (I am also testing models with interaction-effects).
Has anyone ever analysed choice based conjoint data and can tell me if this approach makes sense? In some papers, I read, that they were using random slopes and random intercepts, but I am not sure if this is necessary and how to implement this.
Hopefully, you are able to understand my breadboard and goal. If you have any questions, go ahead and ask!
Thankfully
wlad
I am currently working on a dataset of a forced choice based conjoint experiment and are a little bit troubling on how to analyse the data.
Currently, I am using a mixed-effects model, using the following code:
melogit dependent_var charA1 charA2 charB1 charB2 charC1 charC2 || respondent_ID:
I am currently not sure, if this is correct. The experiment was conducted as follows:
The respondents were presented two objects, that differed in 3 characteristics, and they had to decide, which one they would choose. If they chose object A, this would mark the dependent variable as "1" (made a decision for object A), and simultaneously, the other object that was presented, was marked as dependent_var "0" (did not chose object B). Respondents had to do this 13 times, leading to 26 decisions in total (13 "would chose" decisions, where the dependent variable was then a "1", and 13 "would not chose" decisions, where the dependent variable was then a "0").
The objects differed in three characteristics. These characteristics were always allocated randomly and are coded as dummy variables. This leads to data, that somehow looked as follows:
dependent_var = 1 ; charA1 = 0; charA2 = 1; charB1 = 1; charB2 = 0; charC1 = 1; charC2 = 0; respondent_ID = 45
dependent_var = 0 ; charA1 = 1; charA2 = 0; charB1 = 0; charB2 = 1; charC1 = 0; charC2 = 1; respondent_ID = 45
....(there were 24 more observations for respondent 45, but I cut this here for the sake of brevity).
(This is a little bit reduced, in fact, there were seven characteristics, where some had four different values, and some had three different values).
So these 26 decisions are actually "nested" inside the respondents ID. With logistic regression, I am now trying to find out, which characteristics are the most important ones for the respondents when they want to make a decision. Because of the nested data (I had 300 respondents, each of them made 26 decisions), I want to use a multi-level regression.
Can anyone tell me, if the approach that I described in the beginning, is meaningful? I am also considering to append vce(cluster respondent_ID) to avoid biases according to the small first-level data (i.e., only 26 per respondent), which unfortunately decreases my z-values and makes some coefficients insignificant (I am also testing models with interaction-effects).
Has anyone ever analysed choice based conjoint data and can tell me if this approach makes sense? In some papers, I read, that they were using random slopes and random intercepts, but I am not sure if this is necessary and how to implement this.
Hopefully, you are able to understand my breadboard and goal. If you have any questions, go ahead and ask!
Thankfully
wlad
