1 Two-part model concept and motivation

1.1 Why split outcomes into two components

A two-part model represents an observed outcome with two linked processes. The first process addresses whether a nonzero outcome occurs, while the second process describes the size (intensity) given that the outcome is positive. This separation is useful when zeros are not simply small positive values rounded down, but instead reflect a qualitatively different mechanism (for example, non-participation versus participation, or non-detection versus detected magnitude).

1.2 Typical data patterns (zeros, censoring, heavy tails)

Two-part models are often applied when the data show:

  • Excess zeros relative to standard continuous models (often called zero inflation).
  • Mixtures of a point mass at zero and a continuous distribution over positive values.
  • Right-skewed or heavy-tailed positive values, where the magnitude process differs in scale and variability.

Related challenges include censoring or detection limits, where zeros may arise from the measurement process rather than the underlying intensity.

Two-part models are closely related to other statistical constructions:

  • Hurdle models emphasize a “barrier” at zero: the first part determines whether the hurdle is crossed; the second part models positive outcomes conditional on crossing.
  • Selection-style thinking focuses on how the observation or participation mechanism filters the data, producing a dependence between occurrence and intensity.
  • Mixture perspectives view the outcome as coming from different generating regimes; in two-part form, the regimes are separated into zero vs positive outcomes rather than fully latent mixture classes.

2 Mathematical formulation

2.1 Components of the model

2.1.1 Part 1: occurrence/selection component

Let \(Y\) be an outcome that can be zero or positive. Define an indicator: \[ I = \mathbb{1}(Y>0). \] Part 1 models \(P(Y>0)\) using a binary (or sometimes count-based) regression, such as a logistic or probit model: \[ P(Y>0 \mid x) = p(x;\alpha). \] Here \(x\) denotes predictors and \(\alpha\) denotes parameters for the occurrence process.

2.1.1.1 Mapping latent predictors to probabilities

In logistic regression, a latent index \(x^\top \alpha\) is mapped to probability via the logistic function: \[ p(x;\alpha)=\frac{1}{1+\exp(-x^\top \alpha)}. \] In a probit model, the mapping uses the standard normal cumulative distribution function: \[ p(x;\alpha)=\Phi(x^\top \alpha). \] These links ensure the predicted probability remains in \([0,1]\).

2.1.2 Part 2: intensity/amount component

Part 2 models the distribution of \(Y\) conditional on \(Y>0\). A typical formulation is: \[ Y \mid (Y>0, x) \sim \text{Distribution}(\theta(x;\beta)). \] For strictly positive outcomes, common choices include lognormal, gamma, or Weibull distributions, depending on how the positive tail behaves and on whether the mean scales multiplicatively.

2.1.2.1 Conditional distribution for positive outcomes

For instance, with a gamma model, one might use a mean-structure such as: \[ \mathbb{E}[Y \mid Y>0, x] = \mu(x;\beta), \] with \(\mu(x;\beta)\) linked to predictors through a log link: \[ \mu(x;\beta) = \exp(z^\top \beta). \] Other parameterizations (e.g., modeling the log of \(Y\)) are possible, but the key idea is conditioning on positivity.

2.2 Joint likelihood and factorization

Let \(f_{+}(y\mid x)\) denote the density (or probability mass, if applicable) of \(Y\) for \(y>0\), and let \(p(x)=P(Y>0\mid x)\). Under the two-part construction, the joint likelihood contribution for an observation \((y,x)\) can be written as: \[ L(y\mid x) = \begin{cases} 1-p(x), & y=0,\\ p(x)\, f_{+}(y\mid x), & y>0. \end{cases} \] This factorization expresses the probability of a zero outcome and the conditional positive density as two linked pieces.

2.3 Conditional expectation and marginal effects

A central quantity is the conditional mean of \(Y\): \[ \mathbb{E}[Y\mid x] = P(Y>0\mid x)\,\mathbb{E}[Y\mid Y>0,x]. \] Thus, predictors can influence the overall mean through either component: changing the chance of being positive, changing the expected positive magnitude, or both. Marginal effects on \(\mathbb{E}[Y\mid x]\) often require combining derivatives from each part, with attention to the link functions used.

3 Model variants and naming conventions

3.1 Hurdle models

A hurdle model is commonly described as a two-part model where the first part determines whether the outcome is zero (the “hurdle”), and the second part models the positive values conditional on crossing. In practice, hurdle terminology is frequently used when the probability of zeros is treated as structurally different from the positive process.

3.2 Two-stage (occurrence + positive magnitude) approaches

Two-stage approaches treat the process as sequential: estimate occurrence, then estimate intensity among positive cases. When done separately, this can ignore dependence between parameters across stages, yet it remains popular due to simplicity and interpretability.

3.3 Selection-inspired two-part models

Selection-inspired variants arise when zeros reflect a selection or observation mechanism, implying that the occurrence and intensity processes may be statistically dependent. Some formulations incorporate correction terms or joint modeling to account for such dependence.

3.4 Count-and-continuous hybrids (where applicable)

Although two-part models are often framed for continuous positive outcomes, hybrids are used when the “positive” component is a count (e.g., number of events) or when the outcome blends discrete and continuous regimes in applied settings. In such cases, the model family for part 2 is chosen to match the support and distribution of positive values.

4 Estimation and computation

4.1 Separate vs joint estimation

Two-part models can be estimated:

  • Separately: fit part 1 using all observations (zeros and positives), then fit part 2 using only positive observations. This is computationally straightforward.
  • Jointly: maximize a combined likelihood (or posterior) that includes both components simultaneously. Joint estimation can better capture correlations through shared predictors and coherent uncertainty propagation.

4.2 Maximum likelihood estimation

4.2.1 Deriving score equations (high-level)

Maximum likelihood estimation chooses parameters to maximize the likelihood defined by the two-part factorization. Conceptually, this leads to:

  • A score equation for the occurrence parameters \(\alpha\) based on the Bernoulli-like contributions for zeros and positives.
  • A score equation for the positive parameters \(\beta\) based on the conditional density of positive observations.

When estimated jointly, cross-terms can arise through shared parameters or shared covariates, affecting how standard errors are computed.

4.2.2 Practical implementation details

Implementation typically involves:

  • Selecting link functions for part 1 and part 2 to match data behavior.
  • Ensuring the positive-model density corresponds to the support (strictly positive values).
  • Handling covariates with care (centering/scaling, collinearity checks).
  • Using robust optimization routines and verifying that predicted probabilities and fitted moments are well behaved.

4.3 Bayesian estimation overview

Bayesian estimation uses prior distributions over \(\alpha\) and \(\beta\), producing a posterior distribution for all quantities of interest. This framework naturally supports:

  • Credible intervals for \(\mathbb{E}[Y\mid x]\) and derived marginal effects.
  • Complex hierarchical extensions (see Section 9.1).

In Bayesian two-part models, computation is commonly done via Markov chain Monte Carlo or variational methods, depending on model complexity.

4.4 Dealing with identifiability and convergence

Common computational issues include:

  • Identifiability: if part 1 and part 2 use the same predictors in certain ways and the positive model is weakly informed, multiple parameter combinations may explain the data similarly.
  • Convergence: poor starting values or incompatible distributions for the positive component can cause slow or unstable optimization.

Practical remedies include careful initialization, simplifying overly flexible functional forms, checking scaling, and verifying model support assumptions.

5 Regression modeling and predictors

5.1 Choosing covariates for Part 1 vs Part 2

Covariates can be shared across components or assigned differently based on theory and data behavior. A typical strategy is:

  • Use variables expected to influence participation/occurrence in part 1.
  • Use variables expected to influence magnitude conditional on occurrence in part 2.

Different sets of predictors can improve interpretability, but they also change how the overall mean is driven.

5.2 Interactions and nonlinear terms

Nonlinearity is commonly handled via:

  • Polynomial terms or spline bases.
  • Interaction terms that allow the effect of a predictor on occurrence to differ from its effect on intensity.

Because the overall mean multiplies two components, even simple nonlinearities can produce complex marginal effect patterns that are best summarized with predicted curves or marginal effect calculations.

5.3 Offsets and exposure-style adjustments

When part 2 represents a rate-like quantity or is proportional to a known exposure, offsets can be used. Offsets are often placed in the positive model’s linear predictor. This allows the intensity component to compare expected magnitudes on a standardized scale, while part 1 still captures the chance of a positive outcome.

5.4 Interpreting coefficients across parts

Coefficients in part 1 (e.g., log-odds in logistic regression) reflect changes in the probability of being positive. In part 2, coefficients typically relate to the conditional mean or scale of positive outcomes, depending on the chosen distribution and link. Because the overall mean is the product of these pieces, a predictor that increases \(P(Y>0)\) may also increase \(\mathbb{E}[Y]\) even if its effect on conditional intensity is modest, and vice versa.

6 Model diagnostics and validation

6.1 Goodness-of-fit for occurrence

Diagnostics for part 1 focus on how well the model predicts the binary outcome \(I=\mathbb{1}(Y>0)\). Methods include:

  • Calibration plots for predicted probabilities.
  • Tests or residual diagnostics adapted for binary models.
  • Comparing predicted vs observed positive rates across score bins.

6.2 Goodness-of-fit for positive outcomes

For part 2, diagnostics evaluate distributional adequacy among positive observations. Common tools include:

  • Quantile-quantile plots on the positive scale (or transformed scale, if used).
  • Checks of whether dispersion and skewness match the chosen distribution family.
  • Influence diagnostics to detect observations that strongly affect positive-model parameters.

6.3 Residual and calibration checks

Residuals in two-part settings require care because the data generation differs across zero vs positive cases. Calibration checks are often more actionable than traditional residual sums:

  • Assess whether predicted probabilities and predicted positive means align with empirical summaries.
  • Evaluate whether fitted values reproduce observed patterns such as increasing variance with predictors.

6.4 Predictive performance and cross-validation

Predictive validation can be done via:

  • Cross-validation measuring overall predictive accuracy using scoring rules consistent with the likelihood.
  • Metrics that separate performance on zero vs positive outcomes (e.g., classification metrics for occurrence and likelihood-based metrics for positive magnitude).

This helps identify whether poor performance arises from mis-modeling occurrence, mis-modeling intensity, or both.

7 Inference and interpretation

7.1 Marginal effects on overall mean outcomes

Effects on \(\mathbb{E}[Y\mid x]\) combine contributions from both parts: \[ \mathbb{E}[Y\mid x] = p(x;\alpha)\, m(x;\beta), \] where \(m(x;\beta)=\mathbb{E}[Y\mid Y>0,x]\). Marginal effects can be computed by differentiating this product with respect to predictors. In reporting, it is often clearer to provide changes in predicted mean across realistic covariate values rather than relying solely on raw coefficients.

7.2 Uncertainty quantification (standard errors/credible intervals)

Uncertainty can be summarized using:

  • Standard errors from the observed information matrix in maximum likelihood estimation.
  • Credible intervals from posterior draws in Bayesian analysis.

For derived targets like \(\mathbb{E}[Y\mid x]\) or the probability of positivity, uncertainty should propagate through the nonlinear combination of the two components.

7.3 Sensitivity to distributional assumptions

Because part 2 requires a distributional family, results can be sensitive to the choice of positive distribution and link. Sensitivity checks can include:

  • Comparing alternative distributions for positive outcomes.
  • Re-fitting with different link functions and evaluating whether qualitative conclusions persist.
  • Using flexible distributions or semi-parametric choices in advanced settings (Section 9.3).

8 Applications in applied settings

8.1 Health and utilization outcomes with zero inflation

Health utilization data often contain many individuals with zero recorded use during a follow-up period. Two-part models can distinguish between:

  • The decision or ability to generate utilization (occurrence).
  • The amount of utilization among users (intensity).

This structure supports estimation of how predictors relate to both entering utilization and the subsequent intensity.

8.2 Economics: spending with zeros and positive amounts

Household spending or service expenditures commonly show a large fraction of zeros and a skewed distribution among positive spenders. Two-part models help separate:

  • Whether spending occurs.
  • The conditional magnitude of spending for those who spend.

This is useful for analyzing policy variables, income effects, or preferences without forcing a single continuous model across the entire range.

8.3 Environmental measurements with detection limits

Environmental monitoring may record nondetections as zeros or as values at a detection limit. A two-part model can represent:

  • The probability a measurement is above the reporting threshold (occurrence).
  • The distribution of detected concentrations (intensity).

While specialized censoring models may be preferable when censoring is known precisely, the two-part approach remains a practical baseline.

8.4 Engineering reliability or participation-intensity patterns

In engineering or operations, outcomes can reflect both participation and magnitude, such as:

  • Whether a system undergoes a failure event within a window (occurrence) and the subsequent size measure if it occurs (intensity).
  • Whether users engage with a platform (occurrence) and how intensively they do so (intensity).

Two-part models align well with such “engage or not” plus “how much” structures.

9 Extensions and advanced topics

9.1 Multilevel (hierarchical) two-part models

Hierarchical extensions introduce random effects to capture unobserved heterogeneity across groups (e.g., facilities, regions, cohorts). One may allow random intercepts in part 1, part 2, or both. This can improve borrowing of strength and better reflect correlated responses within groups.

9.2 Time-varying or longitudinal two-part structures

For repeated measures, two-part models can be extended to handle changes over time:

  • Time-varying predictors for occurrence and intensity.
  • Correlation between repeated observations within an individual or unit, implemented via random effects or state-space approaches.

These structures can model dynamic behavior where both the chance of observing a positive value and the conditional magnitude evolve.

9.3 Semiparametric and machine-learning-assisted variants

Advanced variants can relax strict parametric assumptions:

  • Semiparametric modeling of the positive component’s conditional mean function.
  • Flexible learners (e.g., gradient boosting) used to estimate the two components, sometimes with cross-fitting to reduce bias.

These approaches aim to improve fit when the relationship between predictors and positive outcomes is complex, while still preserving the two-part interpretation.

9.4 Handling zero inflation beyond the basic split

When zeros arise from multiple mechanisms, a basic two-part split may be insufficient. Extensions can include:

  • Additional latent classes within the zero region.
  • Models that distinguish structural zeros from sampling zeros.
  • More elaborate mixture structures that retain interpretable decomposition into occurrence and magnitude while acknowledging multiple zero sources.

10 Practical workflow and common pitfalls

10.1 Data preprocessing and defining the two outcomes

A first step is ensuring the modeling target is well defined:

  • Decide what constitutes a “zero” (true zeros, nondetections coded as zero, or counts with structural zeros).
  • Confirm that the positive component truly excludes zeros (positive support consistent with the chosen distribution).

Clear labeling prevents misalignment between the occurrence indicator and the positive-model likelihood.

10.2 Choosing model families for each part

Model family selection should reflect:

  • The distribution of \(I=\mathbb{1}(Y>0)\) for part 1 (often logistic/probit is adequate).
  • The shape of positive values for part 2 (skewness, variance pattern, and scale).

Pragmatically, comparing a small set of candidate positive distributions and checking diagnostics can prevent severe mis-specification.

10.3 Common mistakes (mis-specification, missingness)

Frequent pitfalls include:

  • Using an inappropriate positive distribution (e.g., imposing symmetry when positives are strongly right-skewed).
  • Treating missingness as zeros, which conflates nonresponse with true non-occurrence.
  • Ignoring the distinction between “not observed” and “observed but zero,” especially in measurement settings with detection limits.

10.4 Reporting results in a clear, reproducible way

Transparent reporting typically includes:

  • The occurrence model specification and link function.
  • The positive model family and parameterization (including link and any transformation).
  • How predictors differ (or not) across parts.
  • Presentation of fitted quantities such as predicted probability of positivity and predicted conditional/overall means.

Reproducibility is improved by providing estimation details, convergence settings, and diagnostics used to justify model choices.