Replication: Strange flip of signs in regression

RBOON

New Member
#1
Dear all,

I have successfully run an experiment 2 times.
I predicted an interaction between a dummy variable (intervention 1-yes/0-no) and a metric baseline (T1) affect measure to predict the same affect measure at T2 (post intervention).

The intervention succesfully buffered the effect of the T1 on the T2 measure, so the regression coefficient for the interaction was negative. I used the PROCESS add.in for SPSS to compute the regression.

Now I conducted the same experiment another two times. Again I found a significant interaction in both replications. But this time the signs of the interaction are positive, which doesn´t make any sense at all on a theoretical level.

There are some problems with my data. The criterion at T2 is far off from being normally distributed (pretty skewed with lots of people reporting 0 affect).
Further my predictors are highly correlated and multicollinearity might be a problem.

I suspect some artifiact here but maybe some body has encountered an issue like that before? Any ideas how to deal with this?

Thanks and best regards,
RB