I ran a zero-inflated negative binomial mixed model (ZINB) and now have a statistically significant factor (recording type, either NOCA, BCCH, or Human).
To assess the differences between these three possibilities in the factor, could I rely on a post-hoc tukey HSD? Would this post-hoc test be appropriate for assessing pairwise difference of a factor in a ZINB model?
Also, would I be able to just run separate ZINB mixed models for each pairwise comparison on their own? For example, run one model with a subset of the data that's just NOCA and BCCH, one model that's just BCCH and Human, and one that's just NOCA and Human?
You want to use a method of comparing among groups that is appropriate for the model, and takes into account the whole model. For example, in R, the package emmeans (or lsmeans) can be used for a zero-inflated model specified with the pscl package (see http://rcompanion.org/handbook/J_01.html).
So, you don't want to ignore your model and use a standard Tukey HSD test. Likewise, you don't want to ignore half of your data at a time and use pairwise regression models. Although, I might resort to the latter if I couldn't get software to do the multiple comparisons correctly.
@Ette Etuk, how would you use Dmrt or SNK with a zero-inflated negative binomial model ?
A zero-inflated model shows over-dispersion. The Duncan's multiple range test, for example, proceeds as usual. The treatment means are arranged in order of magnitude. The range of each grouping is compared with the least significant difference (LSD). A range that is larger than the LSD is adjudged significant.
@Ette Etuk,
Duncan mrt ---- along with Tukey HSD and friends ---- assume a conditional normal distribution of the dependent variable. If someone is using a negative binomial, logistic, Poisson, etc. distribution for their regression, obviously they are not treating their data as conditionally normal. And a zero-inflated model assumes that the data are that further from normality.
Minimally, using Dmrt after a negative binomial regression makes no sense. It's like deciding you're going to use screws to put together two pieces of wood, and then using a hammer to pound in the screws.
But also, the p-values are likely to be incorrect from using Duncan mrt on data with a conditional negative binomial, logistic, Poisson, etc., distribution. Why use statistics that you know are likely to be wrong?
In addition, it is silly to use something like Duncan mrt in this case when appropriate methods are readily available from software packages.
Bottom line: Duncan mrt may be an appropriate post-hoc test for one-way anova, but it's not appropriate for negative binomial regression.
Sal Mangiafico So which post hoc test can you use for zero affected or zero inflated models? I am struggling with a GLMM test with zero affected distribution (hurdel model)
Juan Manuel Rojas Ripari , you want to use a method that is appropriate for the model. All of R, SAS, and SPSS have a package or part of a procedure named emmeans. In SAS and SPSS, emmeans is a sub-part of the procedure, so if it's there for that procedure it should be appropriate. In R, emmeans is a separate package, so it will only work reliably with model types it supports. Those are listed here: https://cran.r-project.org/web/packages/emmeans/vignettes/models.html .
I know. I am having trouble in changing the conditional model to de zero affected model in my script
If you are used to R sofware, the lsmeans () or emmeans () functions with 'tukey' adjustment its recommended here. To group the output with letters, don't forget to use the function cld ()
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R