Zero-One Inflated Beta (ZOIB) Model
- Zero-One Inflated Beta (ZOIB) modeling is a three-component framework that integrates discrete point masses at 0 and 1 with a beta distribution for interior values.
- It effectively addresses situations where ordinary beta regression fails due to boundary observations, with practical applications in fields like actuarial science, education, and clinical trials.
- Variations in regression parameterizations, estimation techniques, and computational strategies establish ZOIB as a benchmark for comparing alternative boundary-aware models.
Zero-one inflated beta (ZOIB) modeling denotes a class of mixed continuous-discrete models for responses on the closed unit interval when exact boundary values occur with positive probability and interior values are modeled by a beta distribution. Its defining use case is the failure of ordinary beta regression at $0$ and $1$: the beta law has support only on , so it cannot assign positive probability to exact zeros or ones, while many proportion, rate, and ratio outcomes exhibit point mass at one or both boundaries. Recent arXiv work treats ZOIB both as a standard benchmark for such data and as a baseline against which more specialized boundary-aware models are compared or extended (Chakraborty et al., 2021, Baione et al., 2020, Ye et al., 2022, Solano et al., 26 Jul 2025).
1. Conceptual scope and model family
The central statistical problem addressed by ZOIB is straightforward: some proportion outcomes live on , not merely , and can therefore exhibit exact $0$, exact $1$, and continuously varying interior values. In that setting, the response is neither purely continuous nor purely discrete. ZOIB resolves this by combining discrete mass at the endpoints with a beta density on the open interval, producing a three-part law suited to bounded semicontinuous data (Chakraborty et al., 2021, Baione et al., 2020).
This support structure appears in substantively different domains. In actuarial reimbursement modeling, the three regimes correspond to no reimbursement, full reimbursement, and partial reimbursement (Baione et al., 2020). In school pass-proportion data, the response includes exact zeros and ones at the school level (Chakraborty et al., 2021). In interrupted time series with proportional outcomes, the same decomposition is used to separate zeros, ones, and interior percentages while preserving temporal dependence (Ye et al., 2022). In alcohol use disorder trials, the same geometry is interpreted clinically as complete abstinence, partial drinking, and persistent heavy drinking, although that literature often uses ZOIB primarily as a comparator rather than as the main methodological contribution (Solano et al., 26 Jul 2025).
A recurrent source of confusion is nomenclature. Several nearby model classes are not full ZOIB models even when they are described as “inflated beta.” Some papers fit zero-inflated beta models only, with no mass at $1$ (Fan et al., 2019, Tang et al., 2021, Ahn et al., 6 May 2026). Others study one-boundary inflated beta regression, where or $0$0 is modeled separately but not simultaneously (Ospina et al., 2011, Loose et al., 2015, Queiroz et al., 13 May 2026). This makes “ZOIB” most precise when reserved for the full two-boundary case with mass at both $0$1 and $0$2.
2. Distributional structure and alternative factorizations
A standard ZOIB specification treats the response as a three-component mixture. One explicit formulation writes
$0$3
with
$0$4
so that the full density/mass function is
$0$5
In this parameterization, $0$6 controls total boundary mass, $0$7 allocates that mass between $0$8 and $0$9, and $1$0 parameterize the beta interior through mean and precision (Chakraborty et al., 2021).
An alternative but equivalent-support representation is sequential. In the marginalized ZOIB time-series formulation,
$1$1
with
$1$2
and unconditional density
$1$3
Here $1$4 is the probability of being positive and $1$5 is the conditional probability of being exactly one among positives (Ye et al., 2022).
Recent work makes explicit that the selection order for three-component outcomes is not unique. In the Project MATCH comparison, the BRMS ZOIB benchmark is described as first splitting $1$6 versus $1$7, then splitting $1$8 versus $1$9, whereas the competing HOBZ-BART model uses the alternative order 0 versus 1, followed by 2 versus 3 conditional on 4. The paper states that the selection process is not unique for three-component outcomes and that the alternative ordering induces a different likelihood structure (Solano et al., 26 Jul 2025). This suggests that “ZOIB” refers less to a single canonical factorization than to a family of equivalent-support boundary-inflated beta constructions whose regression interpretation depends on parameterization.
3. Regression parameterizations and interpretive targets
One influential implementation route is the BEINF distribution in GAMLSS. In that formulation, the parameter vector is
5
with reparameterization
6
The paper interprets 7 as the mean parameter of the beta component, 8 as a precision/dispersion-type parameter, 9 as controlling the relative amount of inflation at zero, and 0 as controlling the relative amount of inflation at one. The regression form is
1
with the health-insurance application using
2
3
4
5
The logit link is used for 6 and 7, and the log link for 8 and 9 (Baione et al., 2020).
A different target of interpretation appears in marginalized ZOIB. Instead of regressing the beta-component mean directly, the interrupted time-series model specifies regression on the overall marginal mean
0
The link structure is
1
This parameterization is central because 2 then has direct interpretation on the marginal mean scale rather than only within the interior beta component (Ye et al., 2022).
Bayesian implementations often preserve the same conceptual decomposition while changing the computational machinery. The Project MATCH benchmark describes ZOIB in BRMS/Stan with separate submodels for the boundary and interior parts, weakly informative default priors, HMC/NUTS posterior sampling, and component-specific predictors for zero-inflation, one-inflation, and conditional interior mean. That paper also uses the BRMS-generated predictive expectation
3
as the natural comparator and notes that it is “commonly used in the ZOIB framework” (Solano et al., 26 Jul 2025).
4. Estimation, inference, and computational practice
Maximum-likelihood estimation remains standard in many ZOIB applications. The comparative study of school pass proportions fits ZOIB by maximum likelihood and reports parameter estimates in terms of 4, 5, 6, and 7, using the Kolmogorov–Smirnov statistic to compare ZOIB with inflated unit Lindley alternatives. For the zero-one inflated case, the reported ZOIB K–S values are 8 for aided schools, 9 for government schools, $0$0 for private schools, and $0$1 for all schools combined (Chakraborty et al., 2021). In the actuarial GAMLSS application, likelihood-based fitting is paired with AIC-based link selection, and the logarithmic link for $0$2 and $0$3 is chosen because it yields the smallest AIC (Baione et al., 2020).
Bayesian estimation is common when the regression structure becomes hierarchical or when posterior predictive summaries are of interest. In the Project MATCH benchmark, BRMS/ZOIB is fitted in Stan using Hamiltonian Monte Carlo / No-U-Turn Sampler with weakly informative default priors. The same paper reports substantially longer runtimes for BRMS/ZOIB than for HOBZ-BART, with around $0$4 iterations and $0$5 warmup in the additional simulation discussion (Solano et al., 26 Jul 2025). In the time-series setting, estimation proceeds in two stages: first, maximization of an independence composite likelihood for the marginal parameters; second, pseudo maximum likelihood for the copula dependence parameter $0$6. Standard errors are obtained either from a HAC/Godambe sandwich estimator or from a parametric bootstrap, with the paper recommending bootstrap SEs for small interrupted time series (Ye et al., 2022).
The broader inflated beta literature highlights inferential issues that are directly relevant to ZOIB-type modeling even when the fitted model is one-boundary rather than two-boundary. Small-sample likelihood-ratio inference can be size-distorted, and bootstrap Bartlett correction has been proposed as a practical fix in inflated beta regression (Loose et al., 2015). Likewise, maximum likelihood can be highly sensitive to contamination, and robust estimators plus robust Wald-type tests have been developed for inflated beta regression with a boundary point $0$7 (Queiroz et al., 13 May 2026). Because those papers do not fit simultaneous $0$8- and $0$9-inflation, they are not direct derivations of full ZOIB; however, they indicate that boundary-inflated beta models inherit nontrivial small-sample and robustness issues.
5. Applied domains and substantive interpretation
ZOIB is especially useful when the three components correspond to qualitatively distinct substantive regimes. In health-insurance reimbursement, the response is the ratio between reimbursement and expenditure, and the BEINF model separates episodes with no reimbursement, full reimbursement, and partial reimbursement. In that application, the fitted probabilities of $1$0 and $1$1 are reported to match the observed ones very closely, and the distance for the $1$2 component is reported as less than $1$3 in each branch (Baione et al., 2020).
In school pass-proportion data, ZOIB functions as a benchmark rather than a methodological endpoint. The response lies in $1$4 and includes both exact zeros and exact ones, making ZOIB appropriate in principle. The fitted ZOIB parameter estimates are reported for aided, government, private, and pooled schools, but the paper concludes that the inflated unit Lindley alternative yields smaller K–S distances in every dataset considered (Chakraborty et al., 2021). The significance of this comparison is not that ZOIB is unsuitable, but that it serves as a standard reference model against which alternative closed-interval distributions are assessed.
In interrupted time-series analysis, the model’s value lies in combining boundary inflation with serial dependence and marginal-mean interpretation. The marginalized ZOIB time-series paper applies the model to monthly patient pain-management scores, estimates a change point in October 2010, and reports that neither the level change nor the trend change is statistically significant, while the dispersion parameter increases significantly with $1$5. Because larger $1$6 implies smaller conditional variance, the paper interprets this as reduced variability, with estimated SD dropping from about $1$7 pre-intervention to $1$8 post-change (Ye et al., 2022).
In alcohol use disorder trials, ZOIB appears as the conventional Bayesian comparator for bounded semicontinuous outcomes with point mass at $1$9 and $1$0. The Project MATCH paper argues that the standard ZOIB-style expectation $1$1 is less interpretable for alcohol outcomes because it mixes abstinence, partial response, and persistent heavy use into one summary. In simulations and the applied analysis, HOBZ-BART is reported to outperform BRMS/ZOIB in predictive accuracy, computational efficiency, and PITE estimation; the reported standard deviation of PITEs is around $1$2 under HOBZ-BART versus about $1$3 under ZOIB (Solano et al., 26 Jul 2025). This comparison underscores that ZOIB is often the default benchmark precisely because it is the standard structured model for such data.
6. Limitations, misconceptions, and competing formulations
A common misconception is that any beta-based model for bounded data with endpoint values is a ZOIB model. Recent arXiv work shows a more differentiated landscape.
| Model family | Boundary structure | Example arXiv source |
|---|---|---|
| Full ZOIB | Mass at $1$4, mass at $1$5, beta on $1$6 | (Chakraborty et al., 2021, Ye et al., 2022) |
| Zero-inflated beta | Mass at $1$7, no special mass at $1$8 | (Fan et al., 2019, Tang et al., 2021, Ahn et al., 6 May 2026) |
| Zero-or-one inflated beta | One boundary at a time, $1$9 or 0 | (Ospina et al., 2011, Loose et al., 2015, Queiroz et al., 13 May 2026) |
| Unified closed-interval alternatives | Avoid explicit 1 mixture decomposition | (Hahn, 2023, Kim et al., 16 Sep 2025) |
Traditional ZOIB also has methodological limitations. In the Project MATCH comparison, BRMS/ZOIB is described as relying on a fixed parametric/additive structure and requiring explicit interaction specification, which is burdensome in high-dimensional data with nonlinear effects and interactions. The same paper reports substantially longer runtimes for BRMS/ZOIB than for HOBZ-BART, and notes larger effective parameter counts 2, suggesting higher complexity and potentially more overfitting (Solano et al., 26 Jul 2025). In the SLTB comparison, ZOIB is described as having more parameters and more complexity, and its performance deteriorates in a separation-like setting where the boundary value 3 occurs only in one subgroup; the reported ReadingSkills MSE is 4 for ZOIB versus 5 for SLTB and 6 for XBX, with markedly worse performance on the 7 subset (Kim et al., 16 Sep 2025).
The model class has therefore motivated several alternatives rather than a single replacement. Zero-inflated beta models are used when only 8-inflation matters, as in neural WER prediction for ASR (Fan et al., 2019), ecological percent cover with two distinct zero mechanisms (Tang et al., 2021), and microbiome mediation models that combine a zero-inflation component with a beta mixture for nonzero abundances (Ahn et al., 6 May 2026). Other work seeks to avoid the tripartite decomposition entirely by modeling 9 outcomes with a single closed-interval family, motivated partly by the interpretational and computational complications of augmented beta regression and ZOIB (Hahn, 2023, Kim et al., 16 Sep 2025). A discrete ordinal analogue, inflated discrete beta regression, extends the same mixture intuition to Likert-type scales by placing extra mass at a selected category rather than at continuous endpoints (Taverne et al., 2014).
In that broader context, ZOIB remains the canonical full two-boundary beta-mixture model for $0$00-valued data with excess zeros and ones. Its enduring role is methodological baseline, interpretive scaffold, and reference point: it formalizes the basic three-regime structure of boundary-heavy proportion data, while contemporary work clarifies when that structure is sufficient, when it should be marginalized or extended, and when alternative closed-interval models are preferable (Ye et al., 2022, Solano et al., 26 Jul 2025).