In multi-environment experiments we combine only those locations which show non-significant interactions with the treatments (varieties/genotypes). Could we combine those displaying significant interaction mean squares and how to justify?
Interesting question. From my experience, I cannot think of a reason why pooling two such groups would be advantageous. First, doing so would likely increase variance of the new pooled group. Secondly, I wonder if doing so might interfere with your ability to detect the effects of other independent variables, especially is such a variable interacted with the factor that led to the initial group differences.
Are you attempting to isolate the effects of other factors and you simply want to wash away the variation observed due to another variable? If so, one possibility might be to use a simple random effect of group ID.
A significant interaction does indicate that the environments should be treated singularly. However, If I wanted to screen out some treatments performing the same over the environments and above the mean, would it be justified in such situation or not? It would be good if we could cite a reference to the context.