Zero-Inflated Models
You'll learn how to handle count data with more zeros than standard Poisson or negative binomial models can explain.
The Problem of Excess Zeros
Sometimes when you count events—hospital visits, insurance claims, days absent from work—you observe far more zeros than a standard Poisson or negative binomial model would predict. These aren't just ordinary zeros; they come from two distinct processes:
- Structural zeros: People who would never experience the event (e.g., non-smokers filing smoking-related insurance claims)
- Sampling zeros: People who could experience it but happened not to (e.g., smokers who didn't get sick this year)
Standard count models lump all zeros together and often underestimate them, leading to poor fit and biased predictions.
How Zero-Inflated Models Work
Zero-inflated models combine two sub-models:
Model 1 (Binary): Predicts whether an observation is a "certain zero" (structural) using logistic regression. Think of this as a gatekeeper: "Will this person ever file a claim?"
Model 2 (Count): For those who can have the event, predicts the count using Poisson or negative binomial regression. "Given they might file claims, how many?"
The final prediction blends both: some zeros come from the binary gate being closed, others from the count model landing on zero by chance.
When to Use Them
Use zero-inflated models when:
- Your data has noticeably more zeros than expected
- You can justify two distinct data-generating processes
- Standard Poisson/negative binomial models show poor fit around zero
They're particularly useful in epidemiology, ecology (species counts), and finance (transaction modeling).
Key Takeaway: Zero-inflated models explicitly model excess zeros by combining a binary "certain zero" component with a standard count model, capturing two distinct processes that generate zeros in your data.