Possibly the variable importance is very small in the main effects? that may explain why it is left out in the step wise model.

Another possibility comes from the fact (as Austen and Tu showed) stepwise is unstable(small changes in data give large changes in the variables selected). Most people no longer recommend stepwise for this reason. See the attached for the details and modern alternatives to stepwise, Best wishes, David

http://www4.stat.ncsu.edu/~boos/var.select/lasso.adaptive.html

Hello Ingrid,

What that likely means is that the explanatory information furnished by the variable that was entered at an earlier stage has been made redundant with information furnished by variable(s) entered at a later stage, hence the earlier entered variable no longer contributes significantly to the model. 

As David Booth points out, there are many technical concerns about the use of stepwise methods, not the least of which is, their use does not assure that the best ensemble of variables will be chosen to maximize the collective explanatory or predictive power.  The only way to guarantee that outcome is to try all possible subsets of IVs (e.g., all subsets of 2 IVs, 3 IVs, 4 IVs, etc.), then determine how to trade off parsimony (fewer variables in model) for explanatory power.

Good luck with your work.

Just a quick Comment. Breiman (1995) showed that all subsets isn't stable either. See RF4.pdf above for the cite. About all you have are lasso, bagging, and boosting(see Efron and Hastie,( 2016) also cited in RF4) when you build predictive models.

Thanks your answers are insightfull

When you say one variable is removed - is this one of the levels in your factor? as this might be the base level,,,,

no it is not the base