0
$\begingroup$
  1. Using package "pscl" in R
  2. I have daily count data (i.e. people crossing a river)
  3. My data are overdispersed probably due to excessive zero counts for days
  4. I have categorical (binomial) explanatory variables either 1 or 0

Questions are simply

  1. If you expect zeroes do you use zeroinfl (y~x1+x2+...+xn |y~x1+x2+...+xn, df, ..) in R package pscl? or is this model used only if you want to determine whether the zeroes are true zeroes or otherwise?
$\endgroup$
2
$\begingroup$

Overdispersion and over-abundance of zeros are two different situations that are modeled in different ways.

Overdispersion refers to a variance that increases faster than the mean. For example, a Poisson assumes variance equals mean, and thus increases at the same rate. If variance in fact increases twice as fast as mean does, then the Poisson is not appropriate because of overdispersion. Overdispersion can be picked up in NB and Quasi-Poisson models, for example.

Over-abundance refers to a greater number of zeros than assumed by traditional distributions, like Poisson or Negative Binomial. You can simply fit said distributions to a histogram of your data to see whether you have over-abundance of zeros. You can reflect over-abundance in zero-inflated models, like ZIP and ZINB. You can also consider Hurdle models, but note that the assumptions as to the sources of your zeros differ between Zero-Inflated and Hurdle models.

My suggestion would be to first determine which of these situations is affecting your dataset and proceed accordingly.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Data contains zero values that are not erroneous. The counts I obtain are based on the presence or absence of an individual and abundance of individuals at one specific location from video footage. I am attempting to explain the variance around counts of said individuals as a function of several binomial explanatory (manipulative) variables (e.g. open vs closed; present vs absent) simultaneously. I have confidence that the zero counts are in fact zero and not due to someone falling asleep while watching the video footage. I assuming it is appropriate to use the hurdle model here? Thanks. $\endgroup$ – George Oct 23 '15 at 4:05
  • $\begingroup$ Test both but there is a significant difference in the two types - see stats.stackexchange.com/questions/81457/… $\endgroup$ – Frank H. Oct 23 '15 at 11:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.