Hi, I have peer nomination data collected in three different schools. There are lots of zeros in this data because many youth are not nominated for the items. The data can be thought of as count data, but I need to adjust the variables in some way to account for differences across the schools in the number of nominators. Typically to do so, I create proportion scores (number of nominations received/number of nominators). But, I have a reviewer who is concerned about the large number of zeros and the skew of the data, and is suggesting that I used zero-inflated Poisson or negative binomial models to address my questions (in Mplus, within my SEM framework). My understanding however is that such models can't be run with non-integer data (or proportions) and while they can involve an offset variable, such offset involves log transformations which I can't do because of the large number of zeros. Any advice? Thank you! Julie
Hello Julie,
Yes, when data are counts and possibly wildly skewed, Poisson and negative binomial models are frequently listed as potential go-to choices. There are several options you might consider:
1. Re-express the nomination data as an ordinal variable (e.g., 0, 1, 2, more than 2; or for proportions, something similar). Inspection of the distribution of nominations at the respective schools might suggest possible bins that would not lead to a substantial loss of information from the actual counts.
2. As regards the offset, remember that you can add a constant to all nomination scores, then invoke log transformation. That changes the "zero" point as having indicated a true absence, however.
Good luck with your work!
- Badges
- Science topic
- Similar topics
- Teacher Education
- School