What are the implications of the significance of two variables decreasing when in a model together and increasing when adding an interaction term? In my data set, I have two primary predictors A and B. I also have the outcome Y.
I have created four multivariate models that include other covariates.

*

*Model 1 includes A and covariates,


*Model 2 includes B and covariates,


*Model 3 includes both A and B and covariates, and


*Model 4 includes A and B and covariates and an A**B interaction term.
In Model 1, A is highly associated with Y (Beta=0.037, p<.0001).
In Model 2, B is highly associated with Y (Beta=0.234, P<.0001).
In Model 3, the parameter estimates of both A and B shift closer to 0 and significance decreases (A Beta=0.029, p=.0003; B Beta=0.157, p=.009).
In Model 4, significance and magnitude of both predictors are the highest out of all models (A Beta=0.237, p<.0001; B Beta=0.399, p<.0001; A*B interaction  Beta=-0.066, p<.0001).  .
I also know that A and B are correlated (Spearman's r=0.484, p>.0001). I am looking for assistance in interpretation of these results.
I hypothesize that there may some collinearity in Model 3 without the interaction term.
Does the interaction term serve to diminish that?
In general, I am unsure why significance decreases in Model 3 and increases in Model 4.
 A: In regression there is a danger of omitted-variable bias. In linear regression, if you omit a predictor that is associated both with outcome and with a predictor that is in the model, your results will be biased.
With A and B highly correlated, your results are what you thus might expect. Model 1 attributes to A both its own contribution to outcome and a portion of what B might be contributing on its own, due to the correlation between A and B. Model 2 performs similarly with respect to B.
Model 3 includes both A and B. I suspect that if you look at the variance-covariance matrix of the coefficient estimates (often hidden from view but eventually accessible with a function applied to the model, like vcov() in R), there will be a high-magnitude correlation between their coefficients. I also suspect that predictions based on Model 3 would be superior to those based on either Model 1 or Model 2, even if the individual coefficients for A and B appear to be less "significant."
The significant interaction in Model 4 shows that neither A nor B contribute separately to outcome; the effect of one depends on the level of the other. As the interaction term is associated both with outcome and with each of A and B, its omission from Model 3 would be consistent with omitted-variable bias in Model 3.
One warning on a point that frequently leads to confusion. In Model 4 with the interaction, the individual coefficients for each of A and B represent the situation when the other member of the interaction is at a particular level: 0 for a continuous predictor, or (often) the reference level of a categorical predictor. Re-centering or re-referencing a predictor thus can affect the coefficient of other predictors with which it shares interaction terms.
As the standard test for "significance" is whether a value differs from 0, adding an interaction term can thus make the individual coefficient for one or both of its components appear to be "non-significant." That's an illusion, based on a fairly meaningless type of significance test. When there is an interaction term, proper tests of significance must take into account all of the terms that include a predictor, its interactions as well as its "own" coefficient.
