This article proposes the use of computationally efficient inverse regression Bayesian method for analysis of tweet propagation of political messages. Our example focuses on the Colombian case, though our method can be used in any election where social media messaging has a direct impact on political outcomes. We find strong evidence that politicians were able to identify the combination of sensitive words to enhance the probability of retweet of the message, which, in turn, had an impact on political outcomes. The contributions of our work entail: (a) an examination of a neglected unit of analysis (trigram) in a language less studied (i.e. Spanish), (b) based on an innovative Bayesian efficient approach and (c) exploiting the predictive power that retweets have on electoral results as an informational diffusion tool in social media. A practical implication of this new methodology is the possibility to adjust political messages as a means to increase voters engagement in political campaigns.