Constructing a measure versus using an interaction My dataset is sports data and the outcome variable is wins in a season and I am trying to see the affect of the characteristics of the players on the team on this outcome. My question is, is it easier to use interactions or construct a measure?
I want to know if players attending a top 25 college affects wins when other teamate characteristics are considered. I have 3 possible structures, G1: at least one captain (of the two) went a top 25 and some other or no other players on the team did, G2: no captain attended a top 25 and other players on the team did, and G3: no one on the team attended a top 25 college. I am really interested in comparing G2 to G3 given my outcome. 
Is it best to create 3 variables G1, G2, and G3 and then run my regression with either G2 or G3 as the reference group and see the effect of the coefficient or to crate a dummy for the captains and then the non-captains and then create interactions? I want the most accurate approach, but I am not sure how to setup these interactions or if they will be too confusing to compare to a reference group.
 A: As described in the original post, your independent variables are mutually exclusive: a given team cannot meet more than one of your 3 structures. Thus, there cannot be an interaction term. You will likely be best served with dummy variables for the three structures modeling one as the reference, as you suggest.
If you wanted to evaluate the effect of top 25 schools and the captain/non-captain identity, consider separate variables for captains and other team members and the count of top 25 school attendance of each. With these variables in hand, you could evaluate for interactions between captains and other players (these are not mutually exclusive) or examine for the presence of non-linear effects of top 25 school attendance for either/both.
A: If you want to compare your three groups, just run your regression over them.
If you want to study specifically:


*

*The effect of top25 captain(s) [0 or more]

*The effect of top25 teammate(s) [0 or more]

*The interaction between these two groups


You should study the complete design. In this case you will need a fourth group (no captains, no teammates from top25). The interaction could be hard to observe with small sample or small differences.
