Dear Statalister,
I have cross sectional data in which respondent are asked to rank the importance of 4 features (price, quality, customisation, innovativeness). The data contains 4 variables (price, quality, customisation, innovativeness) and each variable contains numbers from 1 to 4 marking their position, 1 if ranked first, 2 if ranked second, and so on. All respondents rank the 4 characteristics of the same item.
The purpose of the analysis is to "explain" the ranks using a set of respondent characteristics X (age, sex, education, ......).
I could use a multinomial logit if I would limit the analysis to the MOST important feature, only the features which are ranked 1.
However, I would like to understand how the rank of each feature covary with observed respondent's characteristics (would women give more importance to price?) taking into consideration the correlation between the 4 variables.
It is an instance of an ipsative measure: the sum of the 4 variables is always 10, for all respondents. That is, once the 3 out of the 4 features are ranked, the rank of the remaining feature is automatically determined. The four variables are linearly dependent. This introduces a negative correlation between the four variables. Therefore, I was looking for a way to model the 4 variables jointly, as ordered choices. Ideally I would have liked to be able to run something like:
of course, the code does not work of fully ranked data (respondent that actually ranked the 4 features from 1 to 4).
Does anyone have suggestions on how to go about this particular set of variables?
Since there are about 20% of the respondents that do not rank the features perfectly (there are ties), there is some variation in the sum of the ranks of the 4 features. For example, some respondents say that all 4 features are very important (1,1,1,1) some say the all 4 features have little importance (4,4,4,4). When I include these observations I can actually run the cmp model above. However, this is possible only because of the non-perfectly ranked observations. If I were to break the ties the model would not run.
I would like to find a way to correctly model the fully ranked observations and that is also able to handle the partially ranked ones.
My respondents rank only the features for the same object, so I do not have multiple ranks of the same 4 features across various objects/items/vignettes for respondent as in the case of a conjoint analysis.
I do not want to use the ranks given to the 4 features to derive an underlying latent variable as it would be the case using IRT.
Anyone any idea on how to go about these data? Any reference I should be looking into?
Would it make sense to take the difference in rank from a baseline feature? For example, the importance of price is the baseline, then i take difference between the rank of quality and that of price (quality - price, customisation - price, innovativeness - price). The new variables, containing the difference in rank of 3 features from a baseline feature, would then be used in the cmp model. Would this work?
Thanks in advance for your input.
I have cross sectional data in which respondent are asked to rank the importance of 4 features (price, quality, customisation, innovativeness). The data contains 4 variables (price, quality, customisation, innovativeness) and each variable contains numbers from 1 to 4 marking their position, 1 if ranked first, 2 if ranked second, and so on. All respondents rank the 4 characteristics of the same item.
The purpose of the analysis is to "explain" the ranks using a set of respondent characteristics X (age, sex, education, ......).
I could use a multinomial logit if I would limit the analysis to the MOST important feature, only the features which are ranked 1.
However, I would like to understand how the rank of each feature covary with observed respondent's characteristics (would women give more importance to price?) taking into consideration the correlation between the 4 variables.
It is an instance of an ipsative measure: the sum of the 4 variables is always 10, for all respondents. That is, once the 3 out of the 4 features are ranked, the rank of the remaining feature is automatically determined. The four variables are linearly dependent. This introduces a negative correlation between the four variables. Therefore, I was looking for a way to model the 4 variables jointly, as ordered choices. Ideally I would have liked to be able to run something like:
Code:
cmp (price = age, sex, education) (quality = age, sex, education) (customisation = age, sex, education) (innovativeness = age, sex, education), ind(5 5 5 5) cov(unstructured)
Does anyone have suggestions on how to go about this particular set of variables?
Since there are about 20% of the respondents that do not rank the features perfectly (there are ties), there is some variation in the sum of the ranks of the 4 features. For example, some respondents say that all 4 features are very important (1,1,1,1) some say the all 4 features have little importance (4,4,4,4). When I include these observations I can actually run the cmp model above. However, this is possible only because of the non-perfectly ranked observations. If I were to break the ties the model would not run.
I would like to find a way to correctly model the fully ranked observations and that is also able to handle the partially ranked ones.
My respondents rank only the features for the same object, so I do not have multiple ranks of the same 4 features across various objects/items/vignettes for respondent as in the case of a conjoint analysis.
I do not want to use the ranks given to the 4 features to derive an underlying latent variable as it would be the case using IRT.
Anyone any idea on how to go about these data? Any reference I should be looking into?
Would it make sense to take the difference in rank from a baseline feature? For example, the importance of price is the baseline, then i take difference between the rank of quality and that of price (quality - price, customisation - price, innovativeness - price). The new variables, containing the difference in rank of 3 features from a baseline feature, would then be used in the cmp model. Would this work?
Thanks in advance for your input.
Comment