Finding the most relevant factors that lead to an event

I have been tasked with looking at a set of data with about 6000 records that each have 60 or so qualities associated with them (Let these be X1, X2, ... ) and determining what are the top 8 factors that determine whether a record will have a certain designation (there are three, Let's say, A, B, and C). Most of these X's only have two possible values so they are easy to deal with. However there is an X that has 6 values and an X that has 8 values which I have determined to be fairly important. For now, I am just interested in finding the factors amongst the data that lead a record to be event A. My approach so far has been to calculate all of the probabilities of each individual factor (A, B, C and all X's), calculate the probabilities of all of the X's given factor A, and to then use Bayes' theorem to calculate the probability of event A given the X. This gives me a feel for which factors seem to be contributing to event A but I am not sure how I should continue my analysis.

I also delved into looking at intersections of events but I feel like that is a big time sink that yields little to no insight because that requires counting all of the intersections and when you want to look at 5+ factors, the amount of counting you have to do is ridiculous (For example, if you want to look at 5 factors, you need to find 2^5 different events because of the possible ways the events can line up. This causes problems especially when I am looking at the factors that have 6 or 8 different possibilities). I wrote some simple java classes to read in a tab delimited text file of the records and to do all of the counting and calculating for me for individual events and some intersections and I also have an excel spreadsheet that I set up first but that seemed to be fairly slow for performing the calculations. Besides adding specific code (or specific formulas in excel) to pick up each individual event, I am not sure how I can efficiently calculate these probabilities. And I'm not sure if calculating these conditional probabilities of the intersections will help me answer the overall question of what are the top 8 factors that contribute to event A. Any thoughts would be greatly appreciated.

• each one of those 6000 lines records one of three events that occured? What about multinomial logistic regression? Sep 8, 2015 at 20:29

1 Answer

If you believe there is little if any interaction between the independent variables a simple approach is the linear-additive model.

Y= B0 +B1X1 +B2X2 + . . . B60X60 +e

Then by examining the coefficients you can identify the top 8 factors.

This paper may be of some use. If the above model is too simple, there are many other approaches such as neural networks, symbolic regression, splines and others.