First of all, I would like to point out that hierarchical regression, although very appealing, really only helps when one has a relatively sound theoretical framework in mind. Let me elaborate.
It seems that you don’t know what variables should be included in your “final model”, the question like that shouldn’t really be asked in the context of hierarchical analysis as you should include all the possible contributing factors according to a theory, no more and no less. If one includes additional variables one is at the risk of having too many factors controlled for, which may result in a loss of power. If one doesn’t include all the relevant variables, one is at the risk of omitting important explanatory sources, and the conclusions one draws from the analysis might be flawed. Just because some predictors were insignificant in any given analysis doesn’t mean they are irrelevant.
Secondly I would suggest running all of the diagnostics (including outlier analysis) prior to any model testing. The reason is that you want to avoid situations like yours. All uni- bi- and multi- variate outliers should be checked before any data analysis. One finds oneself often in a situation where “removing this one observation could radically transform the outcome”, which is why it is important to do this beforehand.
So my advice would be: start over. Identify any outliers or influential observations, build a model based on theoretically sound ground, and run the analysis with all the relevant variables, without removing any in the “final model”. This is the only way you can really come to any sort of conclusion using hierarchical regression.
Hope this helps!