Two Stage Hurdle Regression for Zero Inflated Data

16 concepts14 papers1 proposal

Imagine trying to predict how much money a patient will spend on a specific medical procedure, the number of fish a commercial vessel will catch on a given day, or the frequency of rare spelling errors in an essay. If you look at the raw data for any of these scenarios, you will notice a striking pattern: a massive, disproportionate number of observations are exactly zero.

Standard statistical models like linear regression or standard Poisson distributions fail spectacularly here. If a single model tries to accommodate the massive spike of zeros, its predictions get dragged downward, systematically underestimating the true magnitude of the non-zero events. Conversely, if the model focuses on capturing the high variance of the positive outcomes, it fails to represent why so many observations remain zero.

To resolve this mathematical tension, researchers across diverse fields decouple the problem into a two-stage architecture known as a hurdle model. The core intuition is simple: an outcome must cross a discrete "gatekeeper" (the hurdle) before its positive magnitude can be realized.

The Anatomy of a Hurdle Model

Instead of forcing a single probability distribution to explain both the absence and the presence of a phenomenon, a hurdle model splits the data generation process into two distinct, independent stages:

  1. Stage 1 (The Hurdle): A binary classification model determines the probability that an observation is non-zero ($Y > 0$ versus $Y = 0$). This acts as a gatekeeper.
  2. Stage 2 (The Intensity): A separate regression model estimates the specific count or continuous magnitude, conditional only on the observation having successfully crossed the hurdle ($Y \mid Y > 0$).

As highlighted by Franzolini et al. in their framework for Bayesian clustering of multiple zero-inflated outcomes, this two-part division provides immense flexibility. Standard zero-inflated models typically assume that zeros can come from two separate sources (e.g., structural zeros and sampling zeros). Hurdle models, by contrast, treat the zero-boundary as a hard threshold, allowing researchers to flexibly model situations with both more zeros or fewer zeros than a standard distribution would expect.

Modeling Continuous vs. Discrete Outcomes

Depending on the nature of the phenomenon being measured, the second stage of the hurdle framework can be adapted to handle either continuous values or discrete counts.

Continuous Magnitudes

When the non-zero outcomes represent continuous values—such as financial costs or physical yields—the second stage typically employs a continuous distribution bounded above zero, such as a lognormal or generalized linear model (GLM).

For example, in healthcare economics, Eddelbuettel et al. utilized this strategy in Changes in Healthcare Spending Attributable to High Deductible Health Plan Offer Among Enrollees with Comorbid Substance Use Disorder and Cardiovascular Disease. The authors used a logit model to first determine whether a patient incurred any healthcare spending at all, and then applied a GLM to estimate the exact dollar magnitude of the expenditure conditional on it being non-zero.

A similar logic applies to ecological forecasting. In Bayesian Modelling of Catch in a North-West Atlantic Fishery, Fernández et al. combined a probit model for zero-catch days with a lognormal regression to forecast positive catches. The authors demonstrated that explicitly separating the zero-occurrence stage prevents the systematic overestimation of yields that occurs when zero-value days are poorly integrated into a single linear model.

When tracking these dynamics over time, the stages can be driven by latent variables. In Visualization of Topic-Sentiment Dynamics in Crowdfunding Projects, do Carmo et al. integrated a two-stage hurdle model into a time-series regression framework, using a latent Gaussian variable to govern the binary probability of a donation occurring, and a log-Normal distribution to model the continuous donation amount once the hurdle was cleared.

Discrete Counts

When the positive outcomes are integers—such as counts of events—the second stage uses a truncated count distribution (like a zero-truncated Poisson or Negative Binomial distribution) that places no probability mass on zero.

In Identifying and analyzing ‘noisy’ spelling errors in a second language corpus, Juffs et al. applied a negative binomial hurdle model to analyze spelling errors made by language learners. This allowed the authors to isolate the factors that make a student prone to making any errors from the factors that dictate the sheer volume of errors they commit once they start making them.

Extending to Non-Parametric and Deep Learning Architectures

The power of the two-stage formulation is that it is entirely modular: you can swap out the underlying models for each stage depending on the complexity of the data.

Tree-Based Hurdle Models

Instead of parametric regressions, the stages can be driven by decision trees to capture complex, non-linear environmental interactions. In Bayesian Classification and Regression Trees for Predicting Incidence of Cryptosporidiosis, Hu et al. paired a classification tree (to map the probability of a disease outbreak's presence) with a subsequent regression tree (to predict the intensity of the outbreak only where it occurred). This allowed the model to uncover distinct environmental "recipes" for disease emergence versus disease magnitude.

Neural Hurdle Networks

For high-dimensional marketing and e-commerce applications, the two-stage split can be embedded directly into deep neural networks. In DNet: Distributional Network for Distributional Individualized Treatment Effects, Wu et al. designed an architecture that pairs a classification head with a distributional regression component. The classifier predicts the probability of a zero outcome (e.g., a customer not buying a product), while the regression component estimates the continuous distribution for non-zero values, allowing for robust predictions even when the underlying data is heavy-tailed.

Input Features (X) ───► [ Shared Representation ]
                             │             │
                             ▼             ▼
                      [Stage 1: Logit]   [Stage 2: Truncated Reg.]
                             │             │
                             ▼             ▼
                       P(Outcome > 0)    E(Magnitude | Outcome > 0)

Overcoming Structural and Causal Challenges

Decoupling occurrence from magnitude does more than just improve predictive accuracy; it solves fundamental structural challenges in causal inference, distributed learning, and missing-data problems.

Causal Discovery in Single-Cell Genomics

In genetics, identifying causal relationships between genes is often stymied by zero-inflation: because most genes in a single cell are "off" at any given moment, single-cell sequencing datasets are dominated by zeros. Traditional causal graph models break down under these conditions. To bypass this, Yu et al. developed Directed Graphical Models and Causal Discovery for Zero-Inflated Data. By modeling the system via a discrete gatekeeper that determines gene activation and a continuous distribution that generates the response magnitude, the authors proved that causal directed acyclic graphs (DAGs) could be accurately recovered without being distorted by the excess zeros.

Privacy-Preserving Distributed Learning

When zero-inflated data is spread across multiple institutions (such as patient records in different hospitals), pooling the data centrally to fit a hurdle model poses severe privacy risks. To solve this, Edmondson et al. developed ODAH (One-shot Distributed Algorithm for Hurdle regression) in An efficient and accurate distributed learning algorithm for modeling multi-site zero-inflated count outcomes. The algorithm separates the problem into a binary classification stage for the zero-threshold and a truncated count model for non-zero values, allowing sites to share only aggregate gradients rather than raw patient data, while still matching the accuracy of a centralized analysis.

Latent States and Missing Labels

In some datasets, zeros are observed not because an event failed to occur, but because the event's magnitude was too small for sensors to detect (e.g., minor wildfires missed by satellites). In Prediction in the presence of response-dependent missing labels, Song et al. addressed this by treating the hurdle framework as a two-level mixture model. By combining a point mass at zero for true non-events with a continuous distribution for event intensities, their framework simultaneously estimates event occurrence and intensity even when the data is only partially observed.

Widespread Application Domains

Because zero-inflation is ubiquitous in real-world data, the two-stage hurdle framework has become a standard tool across a vast array of disciplines:

By recognizing that zero is often a qualitative barrier rather than just a quantitative value, two-stage hurdle models allow researchers to build statistically sound, highly interpretable models that respect the dual-nature of zero-inflated data.

Key Foundational Work

Bayesian Modelling of Catch in a North-West Atlantic Fishery (2002) — contributed 2 concept elements to this direction.
Bayesian Classification and Regression Trees for Predicting Incidence of Cryptosporidiosis (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_two_stage_zero_inflated_modeling_f1be8d96,
  title = {two-stage zero-inflated modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/two-stage-zero-inflated-modeling/rdr_f1be8d961534457bb6684e2403ff96f2},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_two_stage_zero_inflated_modeling_f1be8d96_v1,
  title = {two-stage zero-inflated modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/two-stage-zero-inflated-modeling/rdr_f1be8d961534457bb6684e2403ff96f2/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake