moderated regression analysis

Hi, I am predicting that social inhibition (SI) moderates the relationship between negative affect (NA) and perceived stress. Where higher levels of SI increase the effect of NA on perceived stress. Age and gender were entered on the first step of a multiple regression as control variables as they were correlated with perceived stress, NA and SI on the second step with an SI x NA term entered on the third step. SI and NA were centred before creating the interaction term. all seemed well in the analysis, the control variables accounted for a small but significant amount of stress, on the second step NA accounted for around 50% of the variance and SI for a non significant 2%. however the unstandardised beta weight for the SI term was negative. ( thought it would be positive as the bivariant correlation was about.3 and positive. on the third step the NA x SI term was significant claiming a further 1.5% of the variance of perceived stress and was also negative. when the interaction term was plotted the slope of the lines were in the opposite direction than was anticipated. this suggests that SI moderates between NA and perceived stress where having more SI reduces the effect of NA on stress. this is not only counter intuitive but goes against a substantial amount of previous research. What could be causing this. the study contained 282 participants, assumptions testing proved to be OK, one outlier removed. colinearity statistics maholobis distance 40, Levin's statistic .13. Taking out more participants improved maholobis distance but made no difference in results. NA and SI were correlated by r = .5. I have tested for a supressor variable to no avail.
any thoughts