Timeline for Proper regression model for ratio data
Current License: CC BY-SA 3.0
8 events
| when toggle format | what | by | license | comment | |
|---|---|---|---|---|---|
| Jul 4, 2013 at 15:10 | comment | added | user27588 | @NickCox, thank you, I will certainly take that points. | |
| Jul 4, 2013 at 14:48 | comment | added | Nick Cox | That's going in the wrong direction. Gamma-distributed variables can't be zero, at least not in the standard two-parameter version. I think you need to be more explicit about your concern about "academic support". | |
| Jul 4, 2013 at 14:41 | comment | added | user27588 | What you think about using Gamma GLM since the distibution is heavily skewed? I am mostly concerned about the academic support for the use of Poisson in similar situations. | |
| Jul 3, 2013 at 15:27 | comment | added | Nick Cox | ln(0) is indeterminate, so a ln() transformation is, as already stated, emphatically not an option for you. You can find literature using ln(x + 1) as a work-around, on which there are threads in this forum. I advise against, but there are different views. | |
| Jul 3, 2013 at 15:21 | comment | added | user27588 | The zeroes are sampling zeroes, however the sample is random and I do not expect to be that different from the whole population. I do not believe I will run into unobserved heterogeneity problem. I am still not sure how to proceed, probably I will read more about Poisson. Just to clarify...ln transformation is not an option, right? | |
| Jul 3, 2013 at 14:21 | comment | added | Nick Cox | @Nameless Could well be. But the mean being very low and near zero is consistent with Poisson regression. Also, I would always try Poisson first before trying something more exotic such as a zero-inflated Poisson. | |
| Jul 3, 2013 at 14:18 | comment | added | Nameless | Even if they're all sampling zeros, don't you think he might run into an excess zeros problem using Poisson with 50% or more zeros? | |
| Jul 3, 2013 at 14:00 | history | answered | Nick Cox | CC BY-SA 3.0 |