I hope to get my hands on Michael Lieberman's module on analyzing discrete choice data with SPSS.
I have been caught up with the CBC dilemma for quite some time now and this gives an opportunity to document where I am. Below is an excellent video on running a discrete Choice analysis with SPSS. A proportional hazards model (Cox Regression) is used. I have not worked through the math here but assume this is equivalent to using MNL with the time variables being appropriately defined. The last time I used a Cox regression, t stood for time to a terminal event in a survival analysis. I found the approach interesting because of the design - uses once choice task per respondent (each respondent sees each option once) and the fact that individual characteristics are explicitly modeled (as required by the random utility MNL). Watch the video below:BTW , McFadden's influential 1974 paper is available online at: http://bit.ly/73RmgR. But a much easier read for beginners is the chapter that he references in the video. It is quite detailed and starts with the basics - so if you are already familiar with the basics of logit/logistic regressions, you may skip this. Note that the key to the interpretation is in the odds-ratio which is the exponential of the corresponding beta and you look at whether it exceeds or is less than 1. The interpretation is with respect to the base chosen. The text is available at: http://bit.ly/bC9bbZ . The problem with using SPSS for discrete choice (either MaxDiff or CBC) is that you cannot look at the individual utilities like, for example, you can when running a traditional constant sum Conjoint with SPSS (where you can assess reversals at the individual level). The data would be analyzed in the aggregate. If you are using a Sawtooth Software solution with their Hierarchical Bayes module, this becomes possible. For example, in a MaxDiff, using HB, Sawtooth also returns the individual raw and rescaled scores along with an F-statistic which is the root likelihood parameter times 1000. Essentially, this is done by an assumption on the distribution of the part-worths over all respondents combined with a lower-level estimation as per MNL. This is an efficient way to assess individual cases who may have run through the survey and randomly answered the choice questions (the value of the RLH here would be the same as for a pure chance, random draw) while yet meeting time-stamp criteria that you may have set up for qualifying acceptable data. The catch, however, is that if you want to look at this by individual characteristics - intuitively, you need to segment the data first before using the HB utility estimation. In other words, one needs to allow for differences in the distribution of the part-worths over different segments of the population. The Sawtooth application however uses a simple HB model where all respondents are assumed to come from the same population of individual characteristics. This runs counter to a central piece of random utility models. Note McFadden's warning around this early on in his paper. When individual characteristics are not modeled, these characteristics are, in essence, held to be unobservable/un-measurable/absent. McFadden's approach starts with the definition of an individual with measurable attributes (characteristics) faced with a choice decision. In a paper available from the Sawtooth website, Sentis and Li from Pathfinder Strategies have argued based on their own empirical research, that this does not make a difference to the predictive accuracy of the analysis - predictions on the hold-out were not significantly improved by first segmenting the sample (by demographics and by K-means and Latent class segments) and then applying the HB estimation as compared to running the simple HB model over the entire data. This paper provides little consolation - the data comes from a single dataset. In fact, this can be something that can be picked up by researchers. We seriously need a body of empirical evidence rather than a one dataset study to validate what is definitely a counter-intuitive conclusion. Till such time as such complex HB modeling software does not become available (I am talking about user-friendly software here that allows me to set parameters with a simple GUI), we have no choice but to do one of the following: 1) Look at aggregate level analysis and use logit to directly model interactions between individual characteristics and product attributes multiplicatively - use SPSS plus an Excel simulation tool 2) Use HB estimation, assume that the same size fits all (one population, one distribution), and use averages to get at differences by characteristics 3) Plan for large samples, identify characteristics that are likely to impact choice, separately estimate utilities using HB for each sub-population. Note that unlike a Logit, you won't get statistically defined significance levels for the differences. I would certainly go for 1) or 3).
No comments:
Post a Comment