I am trying to uncover the strength of the association between brand attributes and the likelihood a customer will consider purchasing a certain brand i.e. which attributes are really driving customer consideration.
I have a background in Applied Economics but have not worked with this type of data before, so I am unfamiliar with which method to use.
Data - My independent variables are 6 brand attributes - each respondent has rated them in order of the strength (1 to 10) to which they associate them with 4 different companies (i.e. if attribute A is strongly associated with company B, the datapoint is 10). The dependent variable is the likelihood that the customer would consider that company (1 to 10). We are trying to find a way to remove the company from the analysis and look only to make connections between brand attributes and likelihood of consideration.
I have tried a basic logistic model by lumping the dependent variable into two buckets (likely and not likely), but my dependent variable is not binary and this was useless. We don't necessarily need a regression coefficient but a correlation coefficient or something that can express the relationship (be it causal or not) between attributes and customer consideration.
here is the code I used (i'm sure its amateur) - the event refers to the event that the customer would be likely to consider company x. the data I have actually ranks that likelihood from 1 to 10 so I had to bucket it 1-5 vs. 6-10 which is useless, of course. I need another way to represent the data more fully. For my variables, "twentyfour" refers to the degree to which each respondent associates company x with that characteristic (the company has 24-hour service) - 1 being the weakest and 10 the strongest so, again, this method probably isn't best for that...
proc logistic data=kelly.mark1 outest=kelly.betas covout;
CLASS twentyfour (ref= '1')/PARAM=ref;
CLASS lowerrate (ref= '1')/PARAM=ref;
CLASS service (ref= '1')/PARAM=ref;
class reliable (ref= '1')/PARAM=ref;
class time (ref= '1')/PARAM=ref;
class value (ref= '1')/PARAM=ref;
model likelihood2 (event='1') = twentyfour lowerrate service reliable time value
slentry = .05
output out= kellypred p=phat lower=lcl upper=ucl
predprob = (individual crossvalidate);
** I also ran a contigency table (creating dummy variables for each attribute and rating i.e. twentyfour1, twentyfour2, etc. and running them against likelihood1, likelihood2, etc) - I got contingency coefficients for that and those seem to be useful - but, I am open to other ideas. THANKS!!
You could try a linear model (GLM). If your dependent variable ranges from 1-10, it may have enough spread to allow normal distribution theory to work, especially if the sample size is reasonably large. You can then use the company as a categorical variable and then you can interpret the other regression coefficients as partial correlation coefficients. You might be able to use a MIXED model, as well.
You might also be able to do ordinal logistic regression, but you'll need a lot of data for that model to converge.
You will need to do some reading (Paul Allison's BBUs are good) or get some help from a statistician involved in marketing to get all the pieces and interpretations together. If you are near a major university with a business school, you might be able to get a grad student to help.
If I wanted to use GLM or a mixed model - but my dependent variable was likelihood to consider company x (where values range from 1 to 10) and my independent variables were levels of association each respondent has with x number of product attributes (values range from 1 to 10) ... how would I construct this model to estimate partial effects of each attribute response to the likelihood to consider response. My variables are not discrete, so would I need to use odds ratio estimates and is that possible with GLM or mixed? I still haven't figured this one out, so thanks for any help you can give!