Hi there,

Depending upon the candidate predictors, using the semi-partial r2 could pose some issues. Since relying entirely upon sr2 will remove from consideration any variation that can be explained by more than one of the predictors, you might end up with a less than optimal set of items in the final instrument. Selection will be biased toward including those items that are highly dissimilar to other items yet still associated with the predicted variable.

I feel like the above comment is pretty confusing, so I will try to make it more clear with an example. Lets pretend you are trying to predict income from a set of four variables: years of education, parents SES, spouses SES, and age. Though the first three variables will explain most of the variance in the dependent variable, the final variable (age) may have the largest sr2. This is because the other three are likely to be highly correlated with one another and therefore little of the variance they explain will be unique to each individual predictor. Age, on the other hand, is weakly predictive of income but uncorrelated with the other three variables. Thus, almost all of the relatively modest amount of variance it explains will be reflected in its semi partial r2 since it is unique to this variable.

I can certainly see why some consideration of sr2 is attractive particularly if you are trying to make a brief instrument. Maybe you could try a sort of balanced approach like stepwise (or preferably hierarchical) regression. This could reduce the total number of items you have in the instrument while minimizing the problem detailed above.

Good luck