Description
Models counts where a zero can arise in two ways, because the process never started or because it ran and produced nothing, estimating both parts with their own predictors.
Many behavioural counts have more zeros than a Poisson or negative binomial allows, because some units never engage at all. A zero-inflated model is a mixture: with probability p the process is 'off' and always yields zero; otherwise counts follow a Poisson (or negative binomial) with rate lambda, which can also produce zeros. The likelihood of a zero is p plus (1 - p) times the Poisson probability of zero; a positive count has likelihood (1 - p) times its Poisson probability. Each part gets its own linear model and link, logit for p and log for lambda, and predictors may differ. The model estimates the share of structural zeros even though no individual zero can be classified. Ignoring inflation underestimates the active rate. The two parts can trade off and be weakly identified with little data, so the model should first be validated by simulating data with known values. Hurdle models are a related alternative.
Attributes
- Discipline
- Statistics & Modelling
- Loss type
- Siloed
- Stage
- Post-launch, Evaluation
- Data
- Quantitative
- Effort
- Days
Steps
- State the two processes and why a zero could come from either (never discovered the feature versus discovered it but did not use it this week).
- Simulate data from that story with known values and confirm the model recovers them.
- Fit a zero-inflated Poisson or negative binomial with separate predictors for the zero process (logit) and the count process (log), adding an exposure offset if observation time varies.
- Use weakly informative priors or regularisation if the zero probability is poorly identified.
- Convert both parts to natural scales: probability of a structural zero, and rate when active.
- Compare with plain Poisson, negative binomial and hurdle models using predictive checks of the share of zeros.
Sources
- Statistical Rethinking (McElreath, 2020)
- Finite Mixture Models (McLachlan & Peel, 2000)
- Bayesian Statistical Modelling (Congdon, 2006)
- Univariate Discrete Distributions (Johnson et al., 1992)
- Loss Models: From Data to Decisions (Klugman et al., 2012)
Also known as
- Zero-inflated Poisson
- ZIP
- ZINB
- Zero-inflated negative binomial
- Hurdle-type mixture
Categories
- Analytics & Behaviour
- Statistics & Modelling