Mixture Severity Likelihoods

Why a Mixture Needs Its Own Likelihood

A portfolio may contain losses from two different severity mechanisms. For example, a hypothetical insurer might model smaller equipment claims and larger replacement claims with separate distributions, while recording only each claim amount. The observed loss is complete, but the mechanism that generated it is hidden. That distinction changes the likelihood: each observed amount must allow for either component.

Here, “complete observations” means that every loss value is observed without censoring or truncation. It does not mean the component label is observed. The mixture likelihood sums over the possible labels rather than assigning each loss to whichever component seems more plausible.

Definition: Suppose $X_1,\ldots,X_n$ are independent losses from a two-component severity mixture. Let component 1 have density $f_1(x;\theta_1)$, component 2 have density $f_2(x;\theta_2)$, and let $\pi$ be the probability that a loss comes from component 1, where $0\leq \pi\leq 1$. The density of an observed loss is $$ g(x;\pi,\theta_1,\theta_2)=\pi f_1(x;\theta_1)+(1-\pi)f_2(x;\theta_2). $$ For fully observed losses $x_1,\ldots,x_n$, the likelihood is $$ L(\pi,\theta_1,\theta_2)=\prod_{i=1}^{n}\left[\pi f_1(x_i;\theta_1)+(1-\pi)f_2(x_i;\theta_2)\right]. $$

The component densities must be valid densities on the same observation scale, and the mixing weight must be between zero and one. The formula presumes that the losses are independent and identically distributed from this mixture. If the component label is recorded for each observation, use the labeled-data likelihood instead; if an observation is censored or truncated, its likelihood contribution must also reflect that observation process.

Latent Assignments and Two Likelihoods

One way to describe the hidden labels is to define $Z_i=1$ when loss $i$ comes from component 1 and $Z_i=0$ when it comes from component 2. Each $Z_i$ is Bernoulli with probability $\pi$, independently of the other assignments. If the labels were known, the likelihood would be

L_{\mathrm{labeled}}=\prod_{i=1}^{n} \left[\pi f_1(x_i;\theta_1)\right]^{Z_i} \left[(1-\pi)f_2(x_i;\theta_2)\right]^{1-Z_i}.

But with labels unobserved, an individual loss $x_i$ could have arisen from either component. Adding the two mutually exclusive possibilities gives its observed density:

P(Z_i=1)f_1(x_i;\theta_1)+P(Z_i=0)f_2(x_i;\theta_2) =\pi f_1(x_i;\theta_1)+(1-\pi)f_2(x_i;\theta_2).

Independence across losses then gives the product likelihood in the definition. This is why multiplying the component densities together, or choosing one component for each loss without observed labels, is not the observed-data likelihood.

Worked Example: Constructing an Exponential Mixture Likelihood

A hypothetical severity model uses two exponential components, with rates $\lambda_1=1$ and $\lambda_2=2$ per thousand dollars. Component 1 has mixing weight $\pi=0.4$. Two fully observed losses are $x_1=1$ and $x_2=2$, both measured in thousands of dollars. Find their likelihood under the mixture.

Method. For an exponential distribution with rate $\lambda$, the density is $f(x;\lambda)=\lambda e^{-\lambda x}$ for $x\geq 0$. Use the weighted sum of the two densities for each observation, then multiply the contributions.

At $x_1=1$, the two component densities are

f_1(1)=e^{-1}=0.367879,\qquad f_2(1)=2e^{-2}=0.270671.

Thus the mixture contribution for the first loss is

$$ g(1)=0.4(0.367879)+0.6(0.270671) =0.147152+0.162403=0.309555. $$

At $x_2=2$, the component densities and mixture contribution are

f_1(2)=e^{-2}=0.135335,\qquad f_2(2)=2e^{-4}=0.036631,

$$ g(2)=0.4(0.135335)+0.6(0.036631) =0.054134+0.021979=0.076113. $$

The losses are independent, so the likelihood is

L=g(1)g(2)=(0.309555)(0.076113)\approx 0.02356.

The method avoids the common error of treating the mixture as a choice of one component for the entire sample. Each loss gets its own mixture contribution. The final likelihood is a joint density, not a probability of observing those exact continuous values.

Posterior Probabilities for Hidden Components

Although labels are not observed, the model gives a conditional probability for each possible label after seeing a loss. Define the responsibility $\tau_i$ as the probability that observation $i$ came from component 1, conditional on its observed value and the model parameters. Bayes’ rule gives

\tau_i=P(Z_i=1\mid X_i=x_i) =\frac{\pi f_1(x_i;\theta_1)} {\pi f_1(x_i;\theta_1)+(1-\pi)f_2(x_i;\theta_2)}.

The corresponding probability for component 2 is $1-\tau_i$. A responsibility is a probability, not an observed label: it can be fractional and should not be rounded to a hard assignment when forming the observed likelihood.

Worked Example: Finding Posterior Component Probabilities

Use a hypothetical mixture with exponential component rates $\lambda_1=1$ and $\lambda_2=3$ per thousand dollars, and $\pi=0.5$. Two losses are $x_1=\ln 3$ and $x_2=\ln(3/2)$, in thousands of dollars. Find each loss’s probability of coming from component 1, and the probability that exactly one of the two came from component 1.

Method. Calculate both densities at each loss and apply Bayes’ rule. Because the assignments are independent before observing the losses, they remain conditionally independent given these two separate loss values and fixed parameters.

At $x_1=\ln 3$,

f_1(x_1)=e^{-\ln 3}=\frac{1}{3},\qquad f_2(x_1)=3e^{-3\ln 3}=\frac{1}{9}.

The mixture density there is $0.5(1/3)+0.5(1/9)=2/9$. Therefore,

P(Z_1=1\mid x_1)=\frac{0.5(1/3)}{2/9}=\frac{3}{4}.

At $x_2=\ln(3/2)$,

f_1(x_2)=\frac{2}{3},\qquad f_2(x_2)=3\left(\frac{2}{3}\right)^3=\frac{8}{9}.

The mixture density is $0.5(2/3)+0.5(8/9)=7/9$, so

P(Z_2=1\mid x_2)=\frac{0.5(2/3)}{7/9}=\frac{3}{7}.

Exactly one assignment is component 1 in either of two disjoint cases: the first loss is from component 1 and the second is not, or the reverse. Hence

\frac{3}{4}\left(1-\frac{3}{7}\right) +\left(1-\frac{3}{4}\right)\frac{3}{7} =\frac{3}{7}+\frac{3}{28} =\frac{15}{28}.

The result is about $0.5357$. It avoids the trap of classifying a loss as belonging to the component with the larger density and then treating that classification as certain.

Why the Score Equations Are Difficult

The log-likelihood for unlabelled observations is

\ell(\pi,\theta_1,\theta_2) =\sum_{i=1}^{n}\log\!\left[\pi f_1(x_i;\theta_1)+(1-\pi)f_2(x_i;\theta_2)\right].

For an interior value $0<\pi<1$, differentiation with respect to the mixing weight gives

\frac{\partial\ell}{\partial\pi} =\sum_{i=1}^{n} \frac{f_1(x_i;\theta_1)-f_2(x_i;\theta_2)} {\pi f_1(x_i;\theta_1)+(1-\pi)f_2(x_i;\theta_2)}.

Each summand has a denominator that depends on $\pi$ and both component densities. Differentiating with respect to a component parameter also produces observation-specific weights. For example, if $\theta_1$ appears only in $f_1$,

\frac{\partial\ell}{\partial\theta_1} =\sum_{i=1}^{n}\tau_i \frac{\partial}{\partial\theta_1}\log f_1(x_i;\theta_1).

The weight $\tau_i$ itself depends on the mixing weight and both component models. Thus the component-parameter equations and mixing-weight equation are generally coupled; applying a single-component formula to all losses does not solve them.

Justification of the weighted component score: For one observation, write $g_i=\pi f_1(x_i;\theta_1)+(1-\pi)f_2(x_i;\theta_2)$. If only $f_1$ depends on $\theta_1$, then $$ \frac{\partial}{\partial\theta_1}\log g_i =\frac{\pi\,\partial f_1(x_i;\theta_1)/\partial\theta_1}{g_i} =\frac{\pi f_1(x_i;\theta_1)}{g_i} \frac{\partial}{\partial\theta_1}\log f_1(x_i;\theta_1) =\tau_i\frac{\partial}{\partial\theta_1}\log f_1(x_i;\theta_1). $$ Summing these individual derivatives gives the stated score. The responsibility appears because the component’s density contributes only a fraction of the total mixture density for that observation.

Worked Example: Estimating a Mixing Weight with Fixed Components

Suppose two hypothetical exponential components have fixed rates $\lambda_1=1$ and $\lambda_2=3$ per thousand dollars. The unknown parameter is $\pi$, the weight of component 1. Two fully observed losses are $x_a=\tfrac12\ln 6$ and $x_b=\tfrac12\ln(3/2)$. Find the maximum-likelihood estimate of $\pi$, including a check of the boundaries.

Method. Evaluate the component density ratios, factor out the component 2 density for each loss, and maximize the resulting one-variable likelihood over $0\leq\pi\leq1$.

For an exponential density, $f_1(x)=e^{-x}$ and $f_2(x)=3e^{-3x}$. At $x_a$,

f_1(x_a)=\frac{1}{\sqrt{6}},\qquad f_2(x_a)=\frac{3}{6^{3/2}}=\frac{1}{2\sqrt{6}}, \qquad \frac{f_1(x_a)}{f_2(x_a)}=2.

At $x_b$,

f_1(x_b)=\frac{1}{\sqrt{3/2}},\qquad f_2(x_b)=\frac{3}{(3/2)^{3/2}}, \qquad \frac{f_1(x_b)}{f_2(x_b)}=\frac{1}{2}.

The product of the two component 2 densities is $f_2(x_a)f_2(x_b)=1/3$. Factoring those densities out of the mixture contributions gives

L(\pi)=\frac{1}{3}[1+\pi(2-1)][1+\pi(1/2-1)] =\frac{1}{3}(1+\pi)(1-\pi/2).

Expanding and differentiating,

L(\pi)=\frac{1}{3}\left(1+\frac{\pi}{2}-\frac{\pi^2}{2}\right), \qquad L'(\pi)=\frac{1}{3}\left(\frac{1}{2}-\pi\right).

The stationary point is $\widehat{\pi}=1/2$. Since $L''(\pi)=-1/3<0$, this is a maximum. At the boundaries, $L(0)=1/3$ and $L(1)=1/3$, while $L(1/2)=\frac13(3/2)(3/4)=3/8=0.375$. The interior value is larger, so the maximum-likelihood estimate is $\widehat{\pi}=0.5$. This example avoids the mistaken shortcut of assigning each loss to its more likely component and then estimating the weight from those hard labels.

Practical Checks and Common Traps

Mixture likelihoods can have multiple local maxima and may attain their maximum at a boundary such as $\pi=0$ or $\pi=1$. A score equation identifies stationary points, not automatically the global maximum; the likelihood and feasible boundaries must also be checked. Component labels are interchangeable: swapping the component names and replacing $\pi$ by $1-\pi$ leaves the same mixture density. When interpreting fitted components, an ordering convention can make their labels consistent.

Do not omit mixing weights. The observed density is $\pi f_1+(1-\pi)f_2$, not $f_1+f_2$.
Do not multiply the two component densities for one loss. The loss comes from one component, and its unobserved alternatives are added.
Do not confuse known observations with known labels. Exact loss amounts alone do not reveal their generating components.
Check support and units. Both component densities must be evaluated at each observed loss in the same measurement units.

Mixture models are useful precisely when a population may contain heterogeneous severity patterns but the subgroup membership is unavailable. Their likelihoods preserve that uncertainty instead of forcing unsupported classifications. The central exam skill is to write the weighted sum for each observation first, then form the product across independent observations.

Check Your Understanding

For each problem, use the observed mixture likelihood rather than assuming component labels are known.

A hypothetical mixture has exponential rates $0.5$ and $1.5$ per thousand dollars, with component 1 weight $0.3$. At a loss of $1$ thousand dollars, find the mixture density. (Answer: approximately $0.3253$ per thousand dollars)
Component rates are $2$ and $4$ per thousand dollars, and the component 1 weight is $0.25$. For a loss $x=\ln 2$ thousand dollars, find the posterior probability that it came from component 1. (Answer: $0.4000$)
For two observations, suppose the ratios $f_1(x_i)/f_2(x_i)$ are $3$ and $1/3$, with both component densities fixed and positive. Find the maximizing component 1 weight over $0\leq\pi\leq1$. (Answer: $\widehat{\pi}=0.5$)
Two observed losses have posterior component 1 probabilities $2/3$ and $1/4$. Conditional on the observed losses and fixed parameters, find the probability exactly one came from component 1. (Answer: $7/12$)
A single loss is $0.5$ thousand dollars in a mixture of exponential rates $1$ and $2$ per thousand dollars, with component 1 weight $0.4$. Find its mixture density. (Answer: $0.4e^{-0.5}+0.6(2e^{-1})\approx0.6841$ per thousand dollars)

What You'll Learn