Sophie Leib posted on Tuesday, January 15, 2019 - 12:27 pm
I have count data that is not normal and does not fit a poisson distribution (tested with Kolmogorov-Smirnov). For a mediation analysis, would you recommend: 1. treating the data as count, comparing models (zero-inflated negative binomial, negative binomial) and running the regression with the best fitting model 2. Treating the data as normal and bootstrap the confidence intervals of the outcome?
This depends on the percentage at 0. With a high percentage, perhaps a mixture Poisson (or negative binomial) would work, although a mixture complicates the mediation analysis. Otherwise, a censored-from-below variable might work. With a low percentage, I would treat it as normal and simply use MLR with its robust SEs. I assume it is the Y variable that is of concern.
Sophie Leib posted on Wednesday, January 16, 2019 - 12:39 pm
Thank you for the response. I am treating the data as normal and using MLR estimation. Do you recommend to analyze the standardized or unstandardized indirect effect? Results show different p-values for standardized vs. usstandardized indirect effect. Can you point me to a reference about the difference between the two?
Q1: Either one is fine. We discuss this in our RMA book.
Q2: They have different sampling distributions. Use bootstrapped confidence intervals which allow a nonnormal sampling distribution for the indirect effect by forming a non-symmetric CI. CI's may agree better in terms of covering zero or not.