My scientific interest is to calculate the price elasticity for an overall set of products (books) in a panel dataset of observations over a 3 year period and I was wondering whether asclogit clustering on the individual in this case would be a scientifically (give me the interpretation I want) and computationally (my dataset and computer are up to the task) reasonable approach.
The specifics: I have a panel dataset composed of monthly measurements of whether an individual purchased one or more of 500 book titles while having membership in a book club. The case-specific variables include member ID#, income and education (inferred from addresses) and age. The alternative-specific variables are the book title, book price, and book genre. Choices are not mutually exclusive: any given member on any given month may choose to buy no books, or 1 or more books. This is an unbalanced panel so individuals may not be enrolled for the entire 3 years. Individuals may repeatedly buy the same book in different months.
Each individual for each month they are a club member will have 500 rows representing the 500 book title alternatives that they can buy.
Scientifically, if I want to calculate the mean price elasticity across all books does it make sense to use asclogit with clustering on the individual and then post estimation of the derivatives as below: The STATA the code would be:
asclogit purchased c.price##c.price genre, case(mem_id) alternatives(bk_title) case vars(educ income age) vce(cluster mem_id)
margins, dydx (price)
Can I interpret the results as the effect of a $1 change in price of the theoretical book with average genre on the probability of filling that book for an individual with average education, age and income?
- If this is reasonable scientifically is this reasonable computationally (that is, will the model converge in a reasonable time frame? (I think that the data will fit in memory.)
I have STATA 13.1 SE X64 on Windows 7 with 8GB RAM. The method will essentially be estimating 500 choices (I could potentially drop this list down to the top 200 books if necessary) for 9000 individuals over 3 years. Individuals appear in the dataset on average for 20 months (i.e. they are members for 20 months). There are 7 variables and on average they are 8 bytes each.