I have been thinking about this for a while.
I have a panel dataset with two time periods. My outcome variable is personal income. Since this study was conducted in a low-income country and the entire set of respondents of women, I have a lot of zero, and near-zero values.
I initially changed the zero values into 1 and used the logged form. However, I read that it's not a good practice.
Furthermore, many of those who had zero or near-zero in the first time period had a decent-sized personal income value in the second survey. This leads to a very large percentage change between the two periods (think, from USD 1 to USD 1,000).
This latter problem arises also when I use asinh instead of log.
I know that if I use level value instead of a log, then asymptotically it will take a normal distribution. However, the sample size right now is around 3,000, and I think just using level would lead to inference problems.
How do you tackle this problem? Did you come across anything like this in your case?
