Help Stats :( Interactions and Homoscedasticity

IF anyone could help with this right now would appreciate it so much - have tried looking it up anywhere and finding so many different explanations...

1) When is it appropriate to do an interaction in multiple regression and at what point in my development of the model do I do it... I have already run my model, tested coefficients for significance and started testing assumptions. Can I do an interaction after I have made sure the model doesn't violate assumptios or does it need to be after the creation of dummy variables?

2) Homoscedasticity - have tested this assumption using a residual plot in excel but I have no idea if it is right. WHen i use the regression function it creates a residual plot with one of my independent variables on the x-axis and residuals on the y-axis - but I thought that the x-axis was meant to be the fixed values - so for example 'predicted house prices'.

Also what do I do if I find the residual plot is slightly curved? How do I correct for it easily. I don't want to transform my variables as this will make interpretation harder.

If anyone could help me with these I would be so grateful....feeling so stupid right now :( Thank you so much.
Interaction etc.

Hi Jane,

I think there are basically two reasons to test for interaction of independent variables. First, because you have reason to believe there is interaction from your research question (e.g. a mutual interdependency of your variables was reported before), or second you want to make sure that you have looked into all possibilities before reporting that there is no association of your independent variables with the dependent variable.

In general, i think it doesn't matter much at which point you do the interaction model. If your interaction term is significant there most probably is interaction, if not then not. I guess it is more important to do the interaction model correctly and interprete it carefully.

Homoscedasticity is, however, a basic assumption of the linear regression model, thus it is handy to do this before you do your analyses. In case it turns out that homscedasticity is not met your linear regression model may not be valid, and thus you may not need to do all the linear regression models, but solve the homoscedasticity problem first.