Differences in direction of association in univariate and multivariate analysis


I have a question regarding results from univariate and multivariate analysis.

I have found a positive association between a dependent and an independent variable in a univariate analysis. However, in the multivariate, after controlling for covariates, a negative association has been found. This independent variable is theoretically expected to result in a positive association (as what i've observed in the univariate analysis).

What would a change in direction of association in univariate and multivariate negative binomial regression analysis for count data imply? Would this be an issue of concern? Would this suggest an error? Are there any steps that i can take to check for this?

Help would be greatly appreciated!


Active Member
Yes, this is an issue of concern... One possible explanation is that one of the covariates is strongly co-dependent with the independent variable and takes all the "credit" in the multivariate model. The estimation procedure does not know which of the two variables to give the whole influence weight, since they contain almost the same information.


Less is more. Stay pure. Stay poor.
Question, was it a significant positive then negative predictor. Also, how did the standard error change in magnitude?