Written and maintained by CASRAI Editorial Board
Published
On this page
Beta regression is a regression model for a continuous response that lies strictly between 0 and 1, such as a rate, a proportion or a share. It models the mean of that response through a link function (the logit by default) and describes the spread through a separate precision parameter φ, so the variance changes with the mean instead of being assumed constant.
What is beta regression?
Beta regression was introduced by Ferrari and Cribari-Neto in 2004 under the title “Beta regression for modelling rates and proportions” (Journal of Applied Statistics, 31(7), 799-815). It assumes the response follows a beta distribution and lets explanatory variables enter through the mean. The beta distribution itself is covered in the beta distribution formula guide; beta regression turns its mean into a function of predictors.
The standard reference implementation, the R package betareg, describes it as a model for continuous responses in the open unit interval (0, 1), and gives a reason for preferring it over the older habit of transforming the response and fitting a linear model. Three problems with that habit are listed in the package vignette: the coefficients describe the mean of the transformed response rather than the mean of the proportion itself; proportions usually vary more in the middle of the interval than near 0 or 1, so the constant-variance assumption fails; and their distributions are often asymmetric, which makes Gaussian-based intervals and tests unreliable in small samples. Because the beta variance depends on the mean, beta regression is heteroskedastic by construction.
Beta regression vs other models for proportions
The choice depends on what the response is. The table below separates the cases using the vignette’s rationale and the structure of each model.
| Response | Typical model | Why |
|---|---|---|
| Continuous value in (0, 1), such as a yield, share or index | Beta regression | Models the mean on the logit scale and lets the variance depend on the mean |
| Same, but a transformed response is acceptable | Linear regression on the transformed response | Coefficients then describe the mean of the transformed response, not of y |
| Counts of successes out of a known number of trials | Binomial logistic regression (see logistic regression) | The response is a count with a known denominator, not a continuous proportion |
| Successes out of trials with extra between-unit variation | Beta-binomial (see the beta-binomial distribution) | Adds a beta-distributed success probability to a binomial count |
For the difference between a proportion and a percentage as a response, see proportion vs percentage.
Beta regression formula and symbols
The package vignette writes the beta density in a mean-precision form. If the usual beta shape parameters are p and q, then:
- μ = p / (p + q) is the mean of the response y, so E(y) = μ.
- φ = p + q is the precision parameter. Larger φ means less spread around the mean.
- Var(y) = μ(1 − μ) / (1 + φ).
The mean submodel is g(μᵢ) = xᵢ′β = ηᵢ, where xᵢ is the vector of regressors for observation i, β the coefficient vector, ηᵢ the linear predictor and g the link function. The default link is the logit, g(μ) = log(μ / (1 − μ)), which makes exp(βⱼ) an odds multiplier. The precision can be constant or can have its own submodel, g₂(φᵢ) = zᵢ′γ, with a log link by default in betareg (identity and square-root links are also accepted). Given μ and φ, the shape parameters are p = μφ and q = (1 − μ)φ.
Worked example: gasoline yield
The betareg vignette fits the GasolineYield data: 32 observations of the proportion of crude oil converted to gasoline, with a numeric regressor temp (the temperature in °F at which all gasoline has vaporised) and a ten-level factor batch. The fitted logit model is yield ~ batch + temp. The coefficient for temp is 0.010967 with standard error 0.000413 (z = 26.58), and the precision estimate is φ = 440 (standard error 110). The vignette reports a pseudo R² of 0.962 and a log-likelihood of 84.8 on 12 degrees of freedom.
Odds multiplier. On the logit scale, 10 °F more is a change of 10 × 0.010967 = 0.10967. Then exp(0.10967) = 1.1159, so each additional 10 °F multiplies the odds of the mean yield by about 1.116, holding batch fixed. The 95% Wald interval uses 0.010967 ± 1.96 × 0.000413, which is 0.010158 to 0.011777 per °F. Multiplying by 10 and exponentiating gives 1.107 to 1.125 per 10 °F.
Effect on a proportion. Suppose the mean yield is μ = 0.20 at some temperature. The odds are 0.20 / 0.80 = 0.25. After a 10 °F rise the odds are 0.25 × 1.1159 = 0.2790, so μ = 0.2790 / 1.2790 = 0.218. The same coefficient moves μ by 0.018 here, but by a smaller amount near 0 or 1, because the logit is not linear in μ.
Variance. With φ = 440 and μ = 0.20, Var(y) = 0.20 × 0.80 / 441 = 0.000363, a standard deviation of 0.019. At μ = 0.218 the variance is 0.218 × 0.782 / 441 = 0.000387, a standard deviation of 0.020. The example illustrates the hand arithmetic; the values of μ are chosen for illustration and are not observations from the data set. The data table below lists the odds multipliers for every term in the fitted mean model.
Interpreting beta regression results
| Output | Reading |
|---|---|
| Mean coefficient β (logit link) | Change in the log-odds of the mean response per unit change in the predictor, other terms held fixed |
| exp(β) | Odds multiplier for the mean response; above 1 raises the mean, below 1 lowers it |
| Precision φ (constant) | Larger values mean less variation around the mean; Var(y) = μ(1 − μ)/(1 + φ) |
| Precision submodel coefficient γ | How the spread changes with the variables in the precision formula (log link by default) |
| Pseudo R² | The vignette reports 0.962 for the gasoline model; treat it as a descriptive fit summary, not as a variance-explained share |
| Likelihood-ratio test | Compares nested models; the vignette uses lrtest() from the lmtest package |
Because an effect on the mean depends on where μ sits, report the coefficient and the odds multiplier, and describe changes in μ at stated values of the other predictors. The related effect-size and odds-ratio guidance is in the guide to confidence intervals for an odds ratio.
Assumptions and common mistakes
- The response must lie strictly inside (0, 1). The beta distribution is defined on the open interval. If the response also takes exact 0 or 1, the vignette recommends the transformation (y × (n − 1) + 0.5) / n, citing Smithson and Verkuilen (2006), where n is the sample size. With n = 32 an observed 1 becomes 31.5 / 32 = 0.984375 and an observed 0 becomes 0.5 / 32 = 0.015625. If many values are exactly 0 or 1, that transformation is a weak fix, and the zero or one values deserve their own model.
- Do not use raw residuals. The vignette points out that raw response residuals are unsuitable because the variance depends on the mean. It describes Pearson, deviance and standardized weighted residuals, and notes that
residuals()in betareg has defaulted to quantile residuals since version 3.2-0. - Do not model percentages as if they were counts. A response that is a count of successes out of a known number of trials belongs in a binomial or beta-binomial model, not a beta regression.
- Do not hold φ constant by default without looking. If the spread plainly changes across groups or predictors, fit a precision submodel (
y ~ x1 + x2 | z1 + z2in betareg) and compare with a likelihood-ratio test. - Clustered or repeated measures need a different model. The model described here assumes independent observations; for correlated data, see generalized estimating equations or mixed-effects alternatives.
- Bias correction matters in small samples. The vignette discusses the bias correction of Simas and colleagues (2010), available in betareg as a bias correction or bias reduction option.
How to report beta regression (APA example)
Name the model, the response, the link, the precision treatment and the estimation method, then give each coefficient with its standard error or interval and an interpretable odds multiplier. A sentence in APA style using the vignette’s numbers reads:
A beta regression with a logit link and constant precision was fitted to the proportion of crude oil converted to gasoline (N = 32). Higher end-point temperature was associated with higher yield, b = 0.011, SE = 0.0004, z = 26.58, odds multiplier per 10 °F = 1.12, 95% CI [1.11, 1.12]; estimated precision φ = 440 (SE = 110), pseudo R² = .962.
State the software and package version, any transformation applied to 0 and 1 values, and the residual diagnostics you checked.
How to calculate beta regression in R, Python and Excel
R
library(betareg)
data("GasolineYield", package = "betareg")
fit = betareg(yield ~ batch + temp, data = GasolineYield)
summary(fit)
# variable precision: spread modelled by temp
fit2 = betareg(yield ~ batch + temp | temp, data = GasolineYield)
lmtest::lrtest(fit, fit2)
Python
from statsmodels.othermod.betareg import BetaModel
# df has a response column y strictly between 0 and 1 and a predictor x
model = BetaModel.from_formula("y ~ x", data=df)
result = model.fit()
print(result.summary())
The statsmodels class BetaModel is parameterised by the mean and the precision. Its default link for the mean is the logit and for the precision the log. With arrays rather than a formula it takes exog_precision for the precision variables, and no intercept is added unless you add one, for example with statsmodels.tools.add_constant.
Excel
Excel has no beta regression command, so fit the model in R or Python and use worksheet formulas for the arithmetic on the results. With the temp coefficient in cell B1, the standard error in B2, a starting mean in B3 and φ in B4:
Odds multiplier per 10 units: =EXP(10*B1)
Lower 95% limit: =EXP(10*(B1-1.96*B2))
Upper 95% limit: =EXP(10*(B1+1.96*B2))
New mean after the change: =(B3/(1-B3)*EXP(10*B1))/(1+B3/(1-B3)*EXP(10*B1))
Variance at a mean: =B3*(1-B3)/(1+B4)
With 0.010967, 0.000413, 0.20 and 440, these return 1.1159, 1.1069 to 1.1250, 0.2181 and 0.000363.
Frequently asked questions
What is beta regression used for?
It is used when the outcome is a continuous proportion, rate or index strictly between 0 and 1 and you want to relate its mean to predictors. The package vignette describes it as a model for rates and proportions in the open unit interval.
What does the precision parameter φ mean?
φ is the sum of the two beta shape parameters, p + q. For a fixed mean, a larger φ gives a smaller variance, since Var(y) = μ(1 − μ)/(1 + φ).
Can beta regression handle values of exactly 0 or 1?
Not directly. The model is defined on the open interval (0, 1). The vignette suggests the transformation (y × (n − 1) + 0.5) / n when the response includes 0 or 1, attributing it to Smithson and Verkuilen (2006).
How do I interpret a beta regression coefficient?
Under the default logit link, exp(coefficient) is the multiplicative change in the odds of the mean response per one-unit change in the predictor. In the gasoline example, exp(10 × 0.010967) = 1.116 per additional 10 °F.
Is beta regression the same as logistic regression?
No. Both use a logit link for a mean between 0 and 1, but logistic regression models a binary or binomial-count response, while beta regression models a continuous proportion and adds a precision parameter.
Sources
- Cribari-Neto, F. & Zeileis, A., Beta Regression in R, betareg package vignette, CRAN. https://cran.r-project.org/web/packages/betareg/vignettes/betareg.html (checked 2026-11-15).
- betareg package extensions vignette, CRAN. https://cran.r-project.org/web/packages/betareg/vignettes/betareg-ext.html (checked 2026-11-15).
- Ferrari, S. & Cribari-Neto, F. (2004). Beta regression for modelling rates and proportions. Journal of Applied Statistics, 31(7), 799-815. Record: https://ideas.repec.org/a/taf/japsta/v31y2004i7p799-815.html (checked 2026-11-15).
- statsmodels documentation,
statsmodels.othermod.betareg.BetaModel. https://www.statsmodels.org/stable/generated/statsmodels.othermod.betareg.BetaModel.html (checked 2026-11-15).
Data: Odds multipliers implied by the logit-link beta regression of gasoline yield (GasolineYield data)
Mean-model coefficients and standard errors from the betareg vignette's logit-link fit of yield ~ batch + temp on the GasolineYield data (32 observations). The odds multiplier exp(b) and its 95% Wald interval exp(b ± 1.96 SE) are computed for this page from the published estimates; they are not printed in the vignette.
Period covered: Vignette modified from Cribari-Neto & Zeileis (2010).
| Model term | Coefficient (logit scale) | Standard error | Odds multiplier exp(b) | Lower 95% limit | Upper 95% limit |
|---|---|---|---|---|---|
| temp | 0.010967 | 0.000413 | 1.011 | 1.0102 | 1.0118 |
| batch1 | 1.727729 | 0.101229 | 5.6279 | 4.615 | 6.863 |
| batch2 | 1.322597 | 0.117902 | 3.7532 | 2.9788 | 4.7289 |
| batch3 | 1.57231 | 0.116105 | 4.8178 | 3.8372 | 6.0489 |
| batch4 | 1.059714 | 0.10236 | 2.8855 | 2.361 | 3.5266 |
| batch5 | 1.133752 | 0.103523 | 3.1073 | 2.5367 | 3.8063 |
| batch6 | 1.040162 | 0.106036 | 2.8297 | 2.2987 | 3.4833 |
| batch7 | 0.543692 | 0.109127 | 1.7224 | 1.3907 | 2.1331 |
| batch8 | 0.495901 | 0.108926 | 1.642 | 1.3263 | 2.0328 |
| batch9 | 0.385793 | 0.118593 | 1.4708 | 1.1657 | 1.8557 |
Sources
- Cribari-Neto & Zeileis, Beta Regression in R (betareg package vignette, CRAN) (accessed 2026-11-15)
Compiled by CASRAI from the sources above. Free to reuse under CC BY 4.0 with attribution.







