models are usually applied to data that contains an excess of zeroes compared
to what a classical model can explain (log-normal, Poisson, negative binomial).
To counter the problem, separate models are set up for the probability of
obtaining a positive measurement, and for the positive measurements themselves.
The typical situation is that the model for the
positive numbers allows for either zeroes as natural model results, or values
below a detection limit, but other observations can be zero for completely
different reasons. A standard example is contamination of fish. Some fish may
have a low concentration of a contaminant, below a detection limit, while other
fish may have been living in areas where the contaminant was not present.
However, zero-inflated models often lead to
biased results. There is no clear method to handle bias in zero-inflated
models. The project consists of characterizing the nature of bias in a zero-inflated
model through likelihood approaches, characterize the level of bias from
approaches where a standard model is applied to the positive values (ie. a
standard lognormal, Poisson, or negative binomial), rather than a conditional
model. If the project time allows, solutions can be attempted to be developed.
Knowledge on Likelihood and statistical modelling