Tricky categorical data I am trying to predict damages for accidents with a different number of injuries per accident and each injury could be one of many different types.
For example, one person may have five injuries: an Abrasion Wound to Upper Extremity, a Herniated Lumbar Disk, a Fractured Nose, a chipped canine and a cracked incisor.
The only way I can figure to build the modeling dataset is to create columns such as total number of injuries, total number of disk injuries, total number of dislocations, etc.  Obviously I'll want to group the different injuries up as there are many, many different types; but, other than just using my own intuition and understanding of what should be grouped together I am at a loss of how to do it.
My first thought was mean-encoding or something like that; but, I can only understand how that would work in the case of one variable with many levels.  This situation is new to me.
Anyone who has encountered similar situations, your help would be greatly appreciated. 
 A: In health services research, the Hierarchical Condition Categories risk adjustment model basically predicts some function of your mean expected healthcare spending as a function of a number of disease/healthcare condition dummies, plus age group, plus gender, e.g.
$F[E(Y)] = \beta_0 + \beta_1Diabetes + \beta_2Asthma + \beta_3Heart attack + \beta_4Pneumonia + .....$
The thing is, the model is much more complex than that. They do assign disease groups to hierarchies, and they only count the most severe presentation of a disease within the hierarchy. For example, within the diabetes hierarchy, uncomplicated diabetes is trumped by diabetes with diabetic neuropathy, which is in turn trumped by diabetes with ketoacidosis + coma.
I think you are already proposing some sort of model like that. It does sound like 
a sensible approach. Your problem is that the body is complex, and you're (I assume) not a physician. That said, your example does suggest some groupings. Spinal injuries would clearly be one condition group, and presumably a spinal fracture would trump a herniated disk. Orthodontic injuries could be another group, e.g. fractured tooth trumps chipped tooth.
Overall, an interesting problem, but I don't think this is going to be an easy DIY project. You could make a start with a few groupings that look sensible, though.
