See [Simpson's Paradox][1]. In short the main effect observed can reverse when an interaction is added to a model. At the linked page most of the examples are categorical but there is a figure at the top of the page one could imagine continuously. For example, if you have a categorical and continuous predictor then the continuous predictor could easily flip sign if the categorical one is added and within each category the sign is different than for the overall score. 


  [1]: http://en.wikipedia.org/wiki/Simpson's_paradox