4
$\begingroup$

I am working with presence/absence data that contains lots of zeros. What is the best method to model this. The only suggestions I can find for zero inflated data refer to count data not binomial/proportion data.

Improve this question
$\endgroup$
6
  • 1
    $\begingroup$ I don't understand: Doesn't 0 mean "absence" and 1 "presence"? If not, what are the zeros referring to? $\endgroup$
    – COOLSerdash
    Commented Jun 3, 2013 at 12:28
  • $\begingroup$ Hi yes that is correct. The data specifically is the presence/absence of a certain decision in a game with which there is many rounds. I am therefore using a mixed model with subject id as the random effects. Maybe I am mistaken in thinking that the high number of zeros is a problem in a glmm with binomial errors? $\endgroup$
    – Jonathan Bone
    Commented Jun 3, 2013 at 12:35
  • 1
    $\begingroup$ I've never heard of zero inflated logistic regression (maybe someone else has?). Of course, if you have only zeroes, then you just don't have enough information to fit the model. Does the regression give an error? Something that can happen is complete separation if the outcome variable separates a predictor variable or a combination of predictor variables completely. $\endgroup$
    – COOLSerdash
    Commented Jun 3, 2013 at 12:47
  • 4
    $\begingroup$ The difference between a proportion and a count is the total that you scaled by; surely that can't be a problem; just move between unscaled and scaled counts as needed. If your Bernoulli's are zero-inflated, then as suggested there may be no issue. However, there are zero-inflated binomials in R (VGAM::zibinomial). $\endgroup$
    – Glen_b
    Commented Jun 3, 2013 at 13:26
  • 3
    $\begingroup$ Yes, I think you are mistaken about what is problematic. Zero inflated logistic doesn't make sense to me, and logistic regression makes no assumptions about the proportion who say 0 or 1 $\endgroup$
    – Peter Flom
    Commented Jun 3, 2013 at 13:51

0

Reset to default

Browse other questions tagged or ask your own question.