Background For my master thesis in Public Administration, I'm working on a model to estimate the effect of knowledge claims on retweet count. I'm looking at a data set of 500,000 tweets of EU Parliamentarians. Via a keyword search of words like 'data', 'evidence', or 'statistic', I'm identifying which tweets contain references to knowledge or evidence. I assume that these references are deployed by the political actors to validate their claims and generate greater diffusion of their messages on Twitter.
Problem I've used a multiple linear regression to estimate the effect of the dummy variable 'knowledge claim' (1 for present, 0 for missing) on retweet count, while applying control variables such as follower count and status count. As the results are not very stable across different models, I tried to understand my data better and found that retweets follow a power-law distribution. I believe this is why a linear regression model doesn't produce the best results.
I'm reaching out here to ask if you have a suggestion for me how I can approach this problem. I think if I had a continuous independent variable, I could apply the natural log to both variables to fit the model better, but since my independent variable is dichotomous, I'm not sure if this is the best approach. Should I only apply the log to retweet count? Or do you have any other idea? I'm happy about any ideas or literature suggestions.
I hope this question is not too broad to ask it here.