Modeling Bounded Proportions With Beta Distributions

Why Model a Proportion?

A proportion is naturally bounded: a component’s share of a total, a claim’s fraction assigned to a coverage category, or the fraction of a hypothetical portfolio meeting a condition cannot be less than zero or greater than one. A normal model, which assigns positive probability outside any finite interval, may therefore be unsuitable. The beta distribution provides a flexible continuous model on the unit interval, and its parameters can be chosen to reflect a target average and variability.

A beta model is not automatically appropriate just because a quantity is a proportion. It describes a continuously varying fraction strictly between zero and one, without point masses at the endpoints. The model’s support and its assumptions should be checked against the way the proportion is generated and recorded.

The Beta Model for a Proportion

Definition: A random variable $X$ has a beta distribution with shape parameters $\alpha>0$ and $\beta>0$, written $X\sim\operatorname{Beta}(\alpha,\beta)$, if its density is $$ f_X(x)=\frac{x^{\alpha-1}(1-x)^{\beta-1}}{B(\alpha,\beta)},\qquad 0<x<1, $$ and $f_X(x)=0$ otherwise. Here $B(\alpha,\beta)$ is the beta function. This is a continuous model: $P(X=0)=P(X=1)=0$.

The shape parameters determine how probability is distributed across the interval. For modeling, it is often more direct to describe them using the mean and a concentration parameter. The concentration parameter is the sum of the two shape parameters; it is a way to specify dispersion at a given mean, not necessarily a literal count of observations.

Parameterization: For $X\sim\operatorname{Beta}(\alpha,\beta)$, define the mean $m$ and concentration $\kappa$ by $$ m=\frac{\alpha}{\alpha+\beta},\qquad \kappa=\alpha+\beta. $$ Equivalently, $$ \alpha=m\kappa,\qquad \beta=(1-m)\kappa, $$ where $0<m<1$ and $\kappa>0$. The variance is $$ \operatorname{Var}(X)=\frac{m(1-m)}{\kappa+1}. $$

At a fixed mean, increasing $\kappa$ decreases the variance: the model puts more of its probability near the mean. A smaller concentration gives greater dispersion. The mean determines the balance of the two shape parameters, while the concentration determines their overall scale.

Choosing Parameters From a Mean and Concentration

If a modeling specification gives a target average proportion and a concentration, calculate $\alpha$ and $\beta$ directly. Then use the beta moments to check the implied variance and make sure the parameters are positive. This is useful when the intended level of variability is stated through concentration rather than through a variance.

State the proportion and its scale.
Write $X$ as a fraction between zero and one, not as a percentage between zero and one hundred.

Set the target mean and concentration.
Use $m$ for the mean and $\kappa$ for the sum of the shape parameters.

Calculate and check the shapes.
Set $\alpha=m\kappa$ and $\beta=(1-m)\kappa$. Check that both are positive and sum to $\kappa$.

Worked Example: Matching a Mean and Concentration

Hypothetical model. A modeler represents the fraction of a covered loss assigned to a particular benefit category by $X$. Assume $X$ is continuously distributed on $(0,1)$ and follows a beta distribution. The target mean fraction is $0.40$, and the specified concentration is $20$. Find the shape parameters and the resulting variance.

Method. Convert the mean and concentration into the two beta shape parameters, then verify the implied moments.

The target values are $m=0.40$ and $\kappa=20$. Therefore, $$ \alpha=m\kappa=(0.40)(20)=8, $$ $$ \beta=(1-m)\kappa=(1-0.40)(20)=(0.60)(20)=12. $$ The fitted model is $X\sim\operatorname{Beta}(8,12)$. The shape parameters sum to $8+12=20$, as required.

The implied variance is $$ \operatorname{Var}(X)=\frac{m(1-m)}{\kappa+1} =\frac{(0.40)(0.60)}{20+1} =\frac{0.24}{21} \approx 0.0114286. $$ Thus the standard deviation is $\sqrt{0.0114286}\approx0.1069$. The model’s mean is $8/(8+12)=0.40$, confirming that the parameters reproduce the target mean. The variance is reported to seven decimal places here to show the calculation; it is approximately $0.0114$.

Common mistake avoided. The concentration is $\alpha+\beta$, not either shape parameter alone. Using $\alpha=20$ and $\beta=12$, for example, would change both the mean and the concentration.

Matching a Mean and Variance

Sometimes the target variability is given as a variance rather than as a concentration. For a beta model, the mean and variance determine the concentration and then each shape parameter. The variance must be positive and strictly less than $m(1-m)$ for finite, positive beta shape parameters. If the proposed variance fails this condition, it cannot be matched by an ordinary beta distribution with those moments.

Moment-matching formula: If a beta model is to have mean $m$ and variance $v$, with $0<m<1$ and $0<v<m(1-m)$, then $$ \kappa=\frac{m(1-m)}{v}-1,\qquad \alpha=m\kappa,\qquad \beta=(1-m)\kappa. $$

Why the formula holds. Start with the beta variance expressed using concentration: $$ v=\frac{m(1-m)}{\kappa+1}. $$ Multiply both sides by $\kappa+1$, then divide by $v$: $$ \kappa+1=\frac{m(1-m)}{v}. $$ Subtracting one gives $\kappa=m(1-m)/v-1$. By definition, $\alpha=m\kappa$ and $\beta=(1-m)\kappa$. The condition $v<m(1-m)$ ensures $\kappa>0$; since $0<m<1$, that in turn makes both shape parameters positive.

Worked Example: Recovering Shapes From a Mean and Standard Deviation

Hypothetical model. A modeler represents the fraction of a hypothetical pooled repair cost attributed to one component by $X$. Assume $X$ has a continuous beta distribution on $(0,1)$. The target mean fraction is $0.70$, and the target standard deviation is $0.10$. Find the shape parameters and verify that the target variance is feasible.

Method. Square the given standard deviation to obtain the variance, then apply the moment-matching formula.

First, the target variance is $$ v=(0.10)^2=0.01. $$ The maximum comparison value in the feasibility check is $$ m(1-m)=(0.70)(0.30)=0.21. $$ Since $0<0.01<0.21$, the target variance is feasible. The concentration is $$ \kappa=\frac{m(1-m)}{v}-1 =\frac{0.21}{0.01}-1 =21-1 =20. $$ Now calculate the shape parameters: $$ \alpha=m\kappa=(0.70)(20)=14, $$ $$ \beta=(1-m)\kappa=(0.30)(20)=6. $$ The model is $X\sim\operatorname{Beta}(14,6)$. Check the mean and variance: $$ \frac{\alpha}{\alpha+\beta}=\frac{14}{20}=0.70,\qquad \frac{m(1-m)}{\kappa+1}=\frac{0.21}{21}=0.01. $$ The resulting standard deviation is $\sqrt{0.01}=0.10$, as specified.

Common mistake avoided. The given value $0.10$ is a standard deviation, not a variance. Substituting $0.10$ directly for $v$ would give $\kappa=0.21/0.10-1=1.1$, which does not match the stated standard deviation.

Using the Fitted Model for Probabilities

Once the parameters have been chosen, a probability about the proportion can be found by integrating the beta density over the relevant interval. The beta cumulative distribution function is commonly expressed using the incomplete beta function. For some shape parameters, the density is a simple polynomial and the integral can be evaluated directly.

Worked Example: Probability That a Component Share Is Near One-Half

Hypothetical model. Let $X$ be the fraction of a hypothetical project’s total cost assigned to one of two accounting categories. Assume the fraction varies continuously on $(0,1)$ and follows a beta distribution. The target mean is $0.50$, and the concentration is $4$. Find the probability that the fraction is between $0.25$ and $0.75$, inclusive.

Method. Match the mean and concentration, write the resulting density, and integrate over the requested interval. Because the distribution is continuous, including or excluding either endpoint does not change the probability.

The shape parameters are $$ \alpha=(0.50)(4)=2,\qquad \beta=(1-0.50)(4)=2. $$ Thus $X\sim\operatorname{Beta}(2,2)$. Since $B(2,2)=1/6$, its density is $$ f_X(x)=6x(1-x),\qquad 0<x<1. $$ An antiderivative is $$ \int 6x(1-x)\,dx=\int(6x-6x^2)\,dx=3x^2-2x^3. $$ Therefore, $$ P(0.25\le X\le 0.75) =\left[3x^2-2x^3\right]_{0.25}^{0.75}. $$ At the upper endpoint, $$ 3(0.75)^2-2(0.75)^3 =3(0.5625)-2(0.421875) =1.6875-0.84375 =0.84375. $$ At the lower endpoint, $$ 3(0.25)^2-2(0.25)^3 =3(0.0625)-2(0.015625) =0.1875-0.03125 =0.15625. $$ Subtracting gives $0.84375-0.15625=0.6875$. The probability is $0.6875$ (four decimals). As a check, symmetry of the $\operatorname{Beta}(2,2)$ density about $0.50$ means the two excluded tails have equal probability; each is $0.15625$, so the central probability is $1-2(0.15625)=0.6875$.

Common mistake avoided. The interval probability is the upper cumulative probability minus the lower cumulative probability. Using only the cumulative probability at $0.75$ would include values below $0.25$ as well.

Model Checks and Common Traps

A beta distribution is a model for a continuous fraction, not a universal model for every reported percentage. Before interpreting fitted parameters, check how the data arise and whether the support and variability are compatible with the assumptions.

Check for endpoint observations. A beta variable has zero probability of being exactly zero or one. If a process produces an appreciable number of exact zeros or ones, an ordinary beta model alone cannot represent those point masses. A model with separate endpoint probabilities and a beta distribution for interior values may be more suitable.
Use a fraction, not a percentage, in the formulas. A mean of $40\%$ is $m=0.40$. Substituting $40$ into $\alpha=m\kappa$ would not describe a variable on $(0,1)$.
Do not confuse variance with standard deviation. The moment formula uses $v=\operatorname{Var}(X)$. If a standard deviation is supplied, square it first.
Check feasibility before accepting a fit. For a beta model with a given mean $m$, the variance must satisfy $0<v<m(1-m)$. A calculated nonpositive concentration signals that the proposed moments do not define a valid beta model.
Interpret concentration in context. A larger $\kappa$ means less variance at the same mean. It does not, by itself, prove that the data came from a larger sample or establish how the data were generated.

If the quantity is reported as a percentage $R=100X$, the beta model applies directly to $X=R/100$. Its mean and standard deviation are expressed on the fraction scale; multiplying those quantities by 100 converts them to percentage points. The distributional support, parameter matching, and probability calculations should all be consistent with the chosen scale.

Key takeaway: To model a continuously varying proportion, specify the mean $m$ and either the concentration $\kappa$ or the variance $v$, convert them to $\alpha$ and $\beta$, and verify the implied moments and endpoint assumptions. This parameterization is aligned with the SOA VEE Mathematical Statistics syllabus.

Check Your Understanding

For each problem, treat the stated beta model assumptions as part of the hypothetical scenario. Keep proportions on the zero-to-one scale.

A hypothetical component fraction has target mean $0.30$ and concentration $10$. Find $\alpha$, $\beta$, and the implied variance. (Answer: $\alpha=3$, $\beta=7$, variance $=0.0191$ approximately.)
A hypothetical claim-share fraction has mean $0.60$ and standard deviation $0.11$. Find the concentration and shape parameters, and check feasibility. (Answer: variance $=0.0121<0.24$, $\kappa\approx18.8347$, $\alpha\approx11.3008$, $\beta\approx7.5339$.)
A beta model for a hypothetical fraction has mean $0.25$ and variance $0.015$. Find its concentration and shape parameters. (Answer: $\kappa=11.5$, $\alpha=2.875$, $\beta=8.625$.)
Let a hypothetical interior fraction $X$ follow $\operatorname{Beta}(1,3)$. Find its mean and $P(X>0.50)$. (Answer: mean $=0.25$, $P(X>0.50)=0.1250$.)
A hypothetical allocation fraction is exactly zero for $12\%$ of cases and otherwise varies continuously between zero and one. Is an ordinary beta distribution by itself suitable? (Answer: No; an ordinary beta distribution assigns probability zero to an exact zero, so the model must represent the zero mass separately.)

What You'll Learn