100-Dollar Game Responses as Predictor Outcome = Feature Sentiment
Participants will rate on 1-5 point scale how much like a particular feature.
Potential Predictors = Developments A - E
Participants will be asked to imagine they have ONLY $100 dollars to spend on new developments. We will have a list of 5-15 developments. Participants will then allocate the money to what they want to develop. 
Example Allocation:
Development A: $50
Development B: $30
Development C: $10
Development D: $0 
Development E: $10
We do this across a few participants. I think I can do a  multiple regression to see if the developments that people allocate funds to predict feature sentiment. 
model = lm(Sentiment ~ Development A + Development B + Development C + ...)
It seems simple, but then I realized that each predictor is dependent on the other predictors. That is, the range of values of a predictor varies based on the allocations to other predictors. Do I need to model this dependency? Or, can I get away being, in a sense, naive to these dependencies?  
 A: Whenever you have a degree of freedom that isn't actually a degree of freedom, this can interfere with your model. For example, if $A+B+C+D+E=100$, and my model is $y=\beta_0+\beta_AA+\beta_BB+\beta_CC+\beta_DD+\beta_EE$, you'll observe that if I increase all $\beta_i$ by 1 and $\beta_0$ by -100, I get the same $y$ predictions.
But if my model is instead $y=\beta_0+\beta_AA+\beta_BB+\beta_CC+\beta_DD$, then there isn't the same degeneracy; there should be a unique set of $\beta$ parameters that minimizes the error in my $y$ predictions. (Regularization can also help fix this problem--if I'm trying to minimize the error in $y$ plus the absolute value of the $\beta$ vector, then there's a unique $\beta$ vector that minimizes that combination.)
Most multiple regression techniques will give you some idea about the covariance of the parameter estimates. (They should be negatively correlated--if I overestimate $\beta_A$, that implies I underestimated the other parameters.) They'll also probably handle degeneracy, so you don't need to be too worried about that.
But there's a deeper point to be made about restriction of range--the presence of $A$ can impact how people react to the other options. For example, if $A$ strictly dominates $E$, then I can learn things by comparing how people spend on $A$ and $B$, but will learn the wrong things if I look at how people spend on $B$ and $E$. (I might think that $B$ strictly dominates $E$, for example, when that isn't actually true.)
It's not clear to me the right way to go about this. One approach is to consider combinations (also called cross terms), like $AB$ or $ABC$ or so on. Comparing the AIC of various models (single features, single features plus pairs, everything up to triples, etc.) can tell you if interaction effects are meaningful for your data.
(I would probably try to isolate the pairwise comparisons, asking people to pick $A$ or $B$, and then $B$ or $D$, and so on, by that requires a very different data collection setup and so is likely inapplicable to your problem.)
