Why the Censoring Mechanism Matters
A right-censored claim does not reveal its exact severity. It tells us only that the loss exceeded a known threshold: for example, a claim recorded at a policy limit may represent a larger underlying loss. The usual severity likelihood uses the probability of exceeding that threshold. This is appropriate only under assumptions about how the threshold was set.
The key question is whether the censoring mechanism carries additional information about the underlying loss, beyond the fact that the loss crossed the recorded threshold. If it does, omitting that mechanism can change the likelihood for the severity parameters. This distinction is aligned with the SOA Exam ASTAM (Advanced Short-Term Actuarial Mathematics) syllabus.
For a continuous severity distribution with density \(f_\theta\) and survival function \(S_\theta\), the standard conditional severity likelihood, given censoring thresholds, is
An exact observation contributes a density; a censored observation at \(c\) contributes \(S_\theta(c)=P(X>c)\). This expression is a severity likelihood conditional on the censoring schedule. It does not, by itself, model how that schedule was generated.
When the Censoring Factor Can Be Omitted
Suppose censoring thresholds are generated randomly, independently of losses, and their distribution has parameters \(\eta\) distinct from the severity parameter \(\theta\). Write \(g_\eta\) for the density of \(C\) and \(\bar G_\eta(c)=P(C\ge c)\) for its survival function. For continuous \(X\) and \(C\), the joint observed-data likelihood has the form
where
The censoring factor contains no \(\theta\), so maximizing the joint likelihood over \(\theta\) gives the same severity estimate as maximizing \(L_X(\theta)\). The censoring factor is not literally absent from the joint likelihood; it can be omitted when estimating \(\theta\) because it is a separate factor involving only nuisance parameters.
Record \(t_i\) and whether it is an exact loss or a censoring threshold.
Assess whether censoring is independent of severity, at least conditional on included covariates.
Check that censoring-distribution parameters do not also occur in the severity model.
If either independence or parameter separation fails, the factorization may not support severity estimation.
Why the Factorization Holds
For an uncensored record at \(x\), observing an exact loss \(x\) also tells us that \(C\ge x\). Under independence, the joint contribution is the probability density of the loss times the probability that the censoring threshold reaches it. For a censored record at \(c\), we observe the threshold \(c\) and know \(X>c\). Its contribution is the threshold density times the severity survival probability. Thus
Multiplying these contributions across records groups all severity terms together and all censoring terms together. This is the reason independent censoring permits omission of the censoring factor for estimating \(\theta\). If \(X\) and \(C\) are dependent, the product probabilities need not hold: the distribution of \(C\) among losses of a particular size can differ from its marginal distribution.
Worked Example: A Fixed Independent Censoring Schedule
A hypothetical portfolio has exponential claim severities with density \(f_\lambda(x)=\lambda e^{-\lambda x}\) and survival function \(S_\lambda(x)=e^{-\lambda x}\), for \(x>0\). The observation schedule is fixed independently of severity. One loss is observed exactly at 4, and two losses are censored at thresholds 3 and 5. Find the maximum-likelihood estimate of \(\lambda\). Use loss units consistently throughout.
Method. Condition on the fixed censoring schedule and form the severity likelihood. The exact loss contributes \(f_\lambda(4)=\lambda e^{-4\lambda}\). The two censored losses contribute \(S_\lambda(3)=e^{-3\lambda}\) and \(S_\lambda(5)=e^{-5\lambda}\). Therefore,
The log likelihood is \(\ell(\lambda)=\log\lambda-12\lambda\), so its score is
The second derivative is \(-1/\lambda^2<0\), confirming a maximum. The estimate is \(0.0833\) per loss unit, rounded to four decimal places. Common mistake avoided: the thresholds 3 and 5 are not exact losses, so they contribute survival probabilities, not density factors. The total exposure in the exponential likelihood is \(4+3+5=12\), and there is one exact loss, giving \(\hat\lambda=1/12\).
Worked Example: Keeping the Random Censoring Factor Separate
In a hypothetical study, severities are exponential with rate \(\lambda\), while censoring thresholds are independently exponential with rate \(\mu\). The observed records are one exact loss at 2 and one loss censored at threshold 1. Find the joint likelihood and the maximum-likelihood estimate of \(\lambda\).
Method. For the exact loss, the joint contribution is \(f_\lambda(2)\bar G_\mu(2)=\lambda e^{-2\lambda}e^{-2\mu}\). For the censored record, it is \(g_\mu(1)S_\lambda(1)=\mu e^{-\mu}e^{-\lambda}\). Hence
The severity log likelihood is \(\log\lambda-3\lambda\). Setting its derivative \(1/\lambda-3\) to zero gives \(\hat\lambda=1/3\). The censoring log likelihood is \(\log\mu-3\mu\), giving \(\hat\mu=1/3\). Both second derivatives are negative at their estimates. The severity estimate is \(0.3333\) per loss unit, rounded to four decimal places. Common mistake avoided: omitting the censoring factor is justified for estimating \(\lambda\), not because the observed censoring mechanism has no likelihood contribution, but because its separate factor depends only on \(\mu\).
When Independence or Parameter Separation Fails
Independent censoring is a model assumption, not a consequence of recording a threshold. A common risk factor can affect both severity and the censoring process. For instance, a claim-handling rule might set lower thresholds for a category of claims that also tends to have larger losses. If that category is unobserved or omitted from the model, the censoring threshold can convey information about severity. Conditional independence given recorded covariates can be sufficient, but only if those covariates and their effects are represented appropriately.
Even if the loss and censoring mechanism are independent, the separate-factor argument for estimating \(\theta\) fails when a parameter is shared. For example, if the censoring distribution itself is specified using \(\theta\), its factor is not constant with respect to \(\theta\). In that case the joint likelihood must retain all terms involving \(\theta\).
Worked Example: Severity-Dependent Censoring
Consider hypothetical exponential severities \(X\sim\operatorname{Exp}(\lambda)\). A deliberately informative recording mechanism sets the censoring threshold to \(C=X/2\) for every claim. Two observed records are censored at \(C=1\) and \(C=2\). Estimate \(\lambda\) using the stated mechanism, and compare with the likelihood that would be used under an incorrect independent-censoring assumption.
Method. Since \(C=X/2\), each recorded threshold reveals the exact underlying loss: \(X=2C\). The observed thresholds 1 and 2 therefore correspond to exact severities 2 and 4. The transformation \(C=X/2\), or \(X=2C\), has derivative \(dX/dC=2\). Thus the density of each observed \(C\) is
The correct likelihood for the two observed thresholds is
Its log likelihood is \(\log 4+2\log\lambda-6\lambda\). The score equation \(2/\lambda-6=0\) gives \(\hat\lambda=1/3\); the second derivative \(-2/\lambda^2\) is negative. The estimate is \(0.3333\) per loss unit, rounded to four decimal places.
If one incorrectly treated both observations as ordinary right-censoring at 1 and 2 under independent censoring, the severity likelihood would be \(S_\lambda(1)S_\lambda(2)=e^{-3\lambda}\). This decreases for every \(\lambda>0\), so its supremum occurs as \(\lambda\) approaches the boundary 0, rather than at \(1/3\). Common mistake avoided: a censored threshold is not automatically independent of severity. Here the threshold is a deterministic function of the loss and even reveals the exact loss, so the independent-censoring likelihood discards essential information.
Practical Checks and Exam Traps
Before dropping censoring terms, check both the probability model and the parameterization. The following distinctions are especially useful:
| Situation | Implication for severity likelihood |
|---|---|
| Thresholds are fixed by design, or independent of severity, and parameters are separate | Use the standard severity likelihood; censoring factors are constant in the severity parameters. |
| Independence holds only after conditioning on observed covariates | Use a conditional model that includes those covariates and verify the conditional independence assumption. |
| Censoring depends on severity or an unobserved shared risk factor | The standard factorization may fail; specify the joint observation mechanism. |
| A parameter appears in both severity and censoring distributions | Retain the censoring factor when estimating that parameter. |
A further conceptual trap is to claim that censoring is independent merely because \(C\) was set before the final loss amount was known. The relevant question is statistical: under the assumed model, does knowing the loss change the distribution of the censoring threshold, after conditioning on included information? Also, independence does not mean the observed indicator \(\delta\) is unrelated to the observed loss: large losses can be more likely to exceed a threshold even when \(X\) and \(C\) were generated independently.
Check Your Understanding
For each problem, decide which likelihood contribution is appropriate and check whether the censoring mechanism may be omitted for severity estimation.
- A hypothetical exponential severity model has rate \(\lambda\). Under a fixed independent schedule, losses of 3 and 7 are observed exactly, and losses are censored at 2 and 4. Find the maximum-likelihood estimate of \(\lambda\). (Answer: \(\hat\lambda=2/(3+7+2+4)=1/8=0.1250\) per loss unit.)
- In a continuous severity model, one record is exact at \(x\), and another is censored at \(c\). Write their severity likelihood contribution, conditional on the censoring schedule. (Answer: \(f_\theta(x)S_\theta(c)\).)
- Severities and thresholds are independent exponentials with rates \(\lambda\) and \(\mu\), respectively. The records are exact losses 1 and 3 and censoring thresholds 2 and 4. Find the severity MLE. (Answer: \(\hat\lambda=2/(1+3+2+4)=0.2000\) per loss unit.)
- A mechanism sets \(C=X/3\) for exponential severity with rate \(\lambda\). Two observed censored thresholds are 0.5 and 1.5. Using the mechanism, find the severity MLE. (Answer: The exact losses are 1.5 and 4.5, so \(\hat\lambda=2/(1.5+4.5)=1/3=0.3333\) per loss unit.)
- A censoring distribution is independent of \(X\), but its rate parameter is the same \(\lambda\) used in \(f_\lambda\). Can its likelihood factor be omitted when estimating \(\lambda\)? (Answer: No. Although the mechanisms are independent, the censoring factor depends on \(\lambda\), so it must be retained.)