Partial r is greater than zero-order correlation

I am puzzled by something I have not come across before. To explain, I have conducted a hierarchical multi-regression with a post-therapy outcome measure as criterion and the corresponding pre-therapy outcome measure, alliance measure and relational measure as predictors. This means there are three steps to the model with each single predictor being one block. As well as the usual B, SE, Beta, t, Adj R, I also asked spss for 'part and partial correlations'. In the last (and third) step the partial correlation is greater than the zero-order correlation (the zero-order is -.326 and the partial is -.341). If a partial correlation is the correlation between predictor and criterion after common variance with other predictors and criterion been removed then I would expect it to always be less than the normal zero-order (pearson) correlation. Why is it actually more than the normal (pearson/zero order) correlation in this case? Can anybody throw light on this?


This is probably something to do with suppression. This is when other variables have been artificially reducing the zero-order correlation. What this means is that the true correlation between two of your variables is made small because of the influence of another (or more) variables. In turn the correlation between two of your variables could also be supressing others.

Cohen and Cohen (Applied Multiple Regression/Correlation Analysis) explain this far better than I could in their section on Supression in Regression Models.