Should I compute my scales by averaging or total score? I am conducting a research involving a serial mediation analysis using Hayes's Process (Model 6). Should I compute my scales by average or by total score? It doesn't make a difference considering standardized coefficients, but I'm going to report unstandardized coefficients, which is affected by how you compute your scales (sum or mean).
 A: It doesn't matter. A mean is just a sum divided by the number of items. The substantive conclusion will be the same. The choice depends on what substantively makes more sense. If it's a Likert scale, it probably make more sense to compute an average. If it's a symptom scale, it makes more sense to compute a sum. Note that the fact that this is taking place in a mediation analysis is irrelevant. 
A: A nice way to think of a sum scale is that it is equivalent to adding all the items but constrain the effects tone equal: http://maartenbuis.nl/publications/sum_constr.html If that makes sense for your variables then a sum score is appropriate. 
A: The big question is missing values.  If you have a lot of cases with one or a few indicators missing, either you have to use casewise deletion or allow for missing values.
Let's say that mean.2(X1,X2,X3) and sum.2(X1,X2,X3) will calculate the mean and sum respectively, if there are at least two non-missing values.
Also let's say that you have a number of cases with one of X1,X2,X3 missing, and you'd like to use them.
If you use sum.2(X1,X2,X3) to form a scale, you automatically have lower values for those cases missing an indicator.  If you use mean.2(X1,X2,X3), you don't automatically have that.
Note that you still need to analyze the cases with and without missing values, because not all bias is taken care of (e.g., missing on X1 implies different values of X2 and X3).
