You can split it, but it won't/shouldn't yield different results: The nonsignificant interaction means just that: The influence of the main predictor is the same across all cultures. Now depending on whether "no significant interaction" means something around .8, .2 or .052 the regression slopes for your main predictor will vary more or less. But the overall test tells you, that these differences are not significant, i.e. you have to keep assuming that there are no differences in slopes/regression weights for the main variable of interest in the population.

You can split it, but it won't/shouldn't yield different results: The nonsignificant interaction means just that: The influence of the main predictor is the same across all cultures. Now depending on whether "no significant interaction" means something around .8, .2 or .052 the regression slopes for your main predictor will vary more or less. But the overall test tells you, that these differences are not significant, i.e. you have to keep assuming that there are no differences in slopes/regression weights for the main variable of interest in the population.

Don't split the groups. Instead use dummy variables to distinguish the groups. That way you will get more robust results as you will have a bigger sample.

@frank, markus and Krishna:

Thanks for all your comments and interpretation help. Maybe I also should have given you more details about the variables, yet in general your comments already make sense. I have a total sample size of 241 and for each culture group a minimum of 50ppl. In order to fiind out if my main predictor variable might be moderated through culture/or countries I dummy coded latter variable and computed interaction terms with the main predictor variable and included those in my third regression step. Yet, if I find no main effects of the predictor variable, but moderation effects through culture, then I need to run a simple slope analysis by splitting my data into cultural groups, don't I?

Not unless there's another continuous/intervall scaled moderation effect. With dummy coded variables the interaction weight is the change in slope of the main variable of interest.

Consider the formula:

y= b0 + b1*CUL1 + b2*main_pred + b3*CUL1*main_pred + b4*CUL2 + b5*CUL2*main_pred + ...

if CUL1 is zero, i.e. for everyone not from culture 1, the interaction term behind b3 is also zero. For people from culture 1, CUL1 is 1, and therefore the interaction reduces to b3*main_pred. Which you then can combine with the main effect of main_pred. The new slope is (b2+b3), but only for people in culture 1. Since their CUL2-CULX value is zero, that's all there is. (Plus the covariates of course.) You can also see from this, if b3 is not significant, i.e. probably not different from zero, then the group in question has roughly the same slope as the reference group (b2).

I am not sure what your dependent variable and independent variable. i may be more suitable if you send your regression result.

A difficulty with regression analysis is that it removes the effects one at a time. The next effect in the model may be nearly as large as the first, but you won't know that. Using a true multivariate approach has much less potential for misleading results. I would use a general linear model, or a multiple analysis of variance and covariance using a stat program like SPSS. This would be: MANOVA (dependant variable) by predictor variable (1, # of levels) culture (1, # of categories)/omeans. That analysis will give you the strength of the effect for each of the two factors (indep var, and culture) independent of the other as two possible alternative main effects, and then the interaction of the two on the dependent variable plus the interaction term, and will provide the means to identify which cultures have some differences perhaps. An alternative way is to create dummy variables , (var = euro, 1=euro, 0=not euro),for each culture and use them in the model as covariates. That will identify indivual cultural groups that may affect the dependent varianble when others do not. The MANOVA is robust, and will not miss it if two or more of the variables in the model covary to the extent that they can be "stand-ins" for each other, one almost as strong as the other. At the same time, the multivariate approach protects against type I error, and does not require prior hypotheses about what variables in the model are important and more or less important than others.

Thanks to all of you! highly appreciate your input. Will redo some of my computations and let you know, if i confuse myself again or not in the hopes to receive more of your advise :) cheers