Why the Censoring Mechanism Matters
A right-censored loss record tells us that a loss had not yet been observed by a specified censoring time; it does not reveal the ultimate loss value. The distinction matters because the reason observation stops can affect which losses are seen exactly. Administrative study closure may operate independently of loss size, while a process triggered by the size or development of a loss may censor observations selectively.
The aim here is to diagnose the mechanism, not to infer an unobserved loss from a censored record. A high overall censoring rate does not, by itself, show that censoring is informative. Nor does a low rate establish that it is noninformative. The relationship between losses and the censoring process is the key.
When ties are possible, the convention for assigning an observation with \(X=C\) must be stated. In this tutorial, \(X\le C\) counts as an observed event. This convention does not change the central distinction: a censored record provides a bound on the loss time, not its exact value.
Diagnose the Relationship, Not Just the Rate
A censoring rate is the proportion of records censored, or in a probability model \(P(C<X)\). It is a single summary of the outcome of the censoring process. Independence, by contrast, is a statement about the joint relationship between \(X\) and \(C\): for all relevant sets \(A\) and \(B\), it requires \(P(X\in A,C\in B)=P(X\in A)P(C\in B)\). One probability cannot generally establish all those equalities.
A useful diagnostic, when losses are known or reliably classified, is to compare censoring rates across loss groups. For example, compare \(P(C<X\mid X\text{ is high})\) with \(P(C<X\mid X\text{ is low})\). Different rates are evidence that the censoring event is associated with the loss category. Equal rates across a few categories are not proof of full independence: other relationships may remain, and a limited sample may hide differences.
The following examples use hypothetical data and mechanisms. They separate three questions that are often blurred together: what rule produces censoring, how often censoring occurs, and whether censoring is related to loss size.
Worked Example: A Specified Rule Produces Loss-Dependent Censoring
In a hypothetical portfolio, a loss is classified as high with probability \(0.6000\) and low with probability \(0.4000\). A review threshold is early with probability \(0.5000\) and late with probability \(0.5000\). The joint probability of a high loss and an early threshold is \(0.4000\). Assume explicitly that censoring occurs if and only if the loss is high and the threshold is early. Calculate the censoring rate and compare early-threshold probabilities by loss category.
Method. First use the stated censoring rule to identify the event that counts as censoring. Then compare conditional probabilities. This avoids assuming that a joint probability is a censoring rate without knowing how the mechanism works.
By the rule, the censored event is exactly the intersection of “high loss” and “early threshold.” Therefore,
For a high loss, the probability of an early threshold is
The probability of a low loss and early threshold is the early-threshold probability minus the high-and-early probability:
Thus,
The conditional probabilities differ: \(0.6667\ne0.2500\). The threshold category is associated with loss category, and under the stated rule the censoring rate is \(0.4000\). In particular, the censoring probability is \(0.6667\) for high losses and \(0\) for low losses. This is loss-dependent censoring.
Check. The four joint probabilities are high-and-early \(0.4000\), high-and-late \(0.2000\), low-and-early \(0.1000\), and low-and-late \(0.3000\). They sum to \(1.0000\), and the two censored combinations specified by the rule reduce to the single high-and-early cell, with probability \(0.4000\).
Worked Example: The Same Censoring Rate Under Independence
Consider a different hypothetical portfolio, with loss time \(X\) and censoring time \(C\) independent. Let \(X=2\) with probability \(0.6000\) and \(X=6\) with probability \(0.4000\). Independently, let \(C=1\) with probability \(0.4000\) and \(C=8\) with probability \(0.6000\). Calculate the censoring rate and check whether it varies across the two loss values.
Method. Use independence to form the joint probabilities, then apply the rule \(C<X\) for censoring. This distinguishes independence of the two times from a superficial comparison of overall rates.
If \(X=2\), censoring occurs when \(C=1\), but not when \(C=8\). Therefore \(P(C<X\mid X=2)=0.4000\). If \(X=6\), the same comparison holds: \(1<6\), while \(8>6\), so \(P(C<X\mid X=6)=0.4000\).
The overall probability can also be calculated directly:
The conditional censoring rates are both \(0.4000\), matching the overall rate. The construction also specifies that \(X\) and \(C\) are independent, so this is an independent-censoring model. The calculation illustrates that the same overall rate as in the preceding example can arise under a different dependence structure.
Check. The four joint probabilities for \((X,C)=(2,1),(2,8),(6,1),(6,8)\) are \(0.2400,0.3600,0.1600,0.2400\). They sum to \(1.0000\), and the two censored cells sum to \(0.4000\).
Why One Censoring Rate Cannot Establish Independence
The examples show that an overall censoring rate does not determine whether censoring is loss-dependent. In the first example, the rate \(0.4000\) arises from a rule that targets a particular loss-threshold combination. In the second, the rate is also \(0.4000\), but \(X\) and \(C\) are independent. The rate alone is identical; the mechanisms are not.
The reason can be stated using a basic independence check. If \(A\) denotes “loss is high” and \(B\) denotes “threshold is early,” independence requires \(P(A\cap B)=P(A)P(B)\). In the first example, the product of the marginal probabilities is \((0.6000)(0.5000)=0.3000\), but the joint probability is \(0.4000\). Since \(0.4000\ne0.3000\), the categories are dependent.
This check also clarifies why a censoring rate alone is insufficient. The censoring rate describes the probability of a particular event determined by a rule involving the loss and censoring process. It does not provide all the joint probabilities needed to test independence. Even knowing the overall rate and the loss-group proportions may not suffice unless the censoring rule or the group-specific censoring outcomes are also known.
Worked Example: Comparing Censoring Rates in a Validation Sample
A hypothetical validation study has reliable ultimate-loss values for 24 claims, including claims that would ordinarily be censored. Eight losses are classified as high and 16 as low. The censoring process would censor 6 of the 8 high losses and 4 of the 16 low losses. Calculate the overall and group-specific rates and interpret the evidence.
Method. Use the validation information to calculate rates within actual ultimate-loss groups. This is possible here because the study provides losses even for records that would otherwise be censored; ordinary right-censored records generally do not provide that information.
The overall number censored is \(6+4=10\), out of \(24\). Hence the overall rate is
Among high losses, the rate is \(6/8=0.7500\). Among low losses, it is \(4/16=0.2500\). Thus the high-loss rate exceeds the low-loss rate by \(0.7500-0.2500=0.5000\), or 50 percentage points. In this validation sample, censoring is strongly associated with loss category.
Check. Weighting the two group rates by their group sizes gives \((8/24)(0.7500)+(16/24)(0.2500)=0.2500+0.1667=0.4167\), which agrees, subject to rounding, with \(10/24\).
Common mistake avoided. Treating \(0.4167\), the overall rate, as evidence of noninformative censoring would ignore the markedly different group rates. Conversely, this sample association is evidence, not a mathematical proof of the full population dependence structure.
What Can Be Diagnosed From Ordinary Records?
In routine right-censored data, \(X\) is unknown whenever \(\Delta=0\). Therefore, directly comparing censoring rates across categories of the ultimate loss may be impossible: the censored claims do not reveal which ultimate-loss category they belong to. This is a limitation of the observed data, not evidence that censoring is independent.
A practical assessment should begin with the process that generated the censoring. Was observation stopped at a fixed administrative date set independently of individual losses? Did a claim-management or reporting rule stop follow-up because of information related to claim size? Were different policy, operational, or study conditions applied to different groups? These questions can identify plausible dependence mechanisms and guide what can be checked.
If relevant variables are observed for everyone, compare censoring patterns across those variables. A difference can show that censoring depends on an observed factor. It does not automatically show that censoring depends on the unobserved ultimate loss after accounting for that factor. Conversely, similar observed rates do not rule out dependence on unobserved loss information. Independence conditional on observed covariates is a stronger modeling assumption than merely finding similar overall rates.
The distinction has practical consequences for empirical loss and survival summaries. If censoring preferentially removes losses with particular characteristics, the observed exact losses may not represent the full loss distribution. The direction and size of any resulting distortion depend on the mechanism; they cannot be determined from the censoring rate alone.
Common Mistakes and Exam Traps
- Equating a joint probability with the censoring rate without stating the rule. In Example 1, \(0.4000\) is the censoring rate only because censoring was explicitly defined to occur if and only if the loss is high and the threshold is early. If censoring instead occurred whenever the threshold was early, the rate would be \(0.5000\). The probabilities alone do not specify which rule applies.
- Assuming equal overall rates imply equal mechanisms. Both first examples have a \(0.4000\) censoring rate, but one is loss-dependent and the other is constructed with independent loss and censoring times.
- Calling a conditional difference proof of every kind of dependence. Different group-specific rates are evidence of association for the categories examined. They do not describe every possible relationship or establish a complete model.
- Using an unknown ultimate loss as though it were recorded. When \(\Delta=0\), the exact \(X\) is unavailable. Do not classify that record by its unobserved ultimate loss unless a validation source or other reliable information supplies it.
- Interpreting no observed difference as proof of independence. Small samples, coarse categories, or unobserved factors can conceal dependence.
Check Your Understanding
For each problem, identify the relevant censoring rule before interpreting the rate.
- A hypothetical portfolio has a high-loss probability of \(0.4000\), an early-review probability of \(0.3000\), and a high-and-early probability of \(0.1800\). Censoring occurs if and only if a loss is high and review is early. Find the censoring rate and \(P(\text{early}\mid\text{high})\). (Answer: Censoring rate \(0.1800\); conditional probability \(0.1800/0.4000=0.4500\).)
- In a hypothetical study, \(X\) and \(C\) are independent. \(X\) is 3 or 7 with equal probabilities; \(C\) is 1 with probability \(0.2500\) and 9 with probability \(0.7500\). Censoring occurs when \(C<X\). Find the censoring rate and each conditional censoring probability given \(X\). (Answer: Overall rate \(0.2500\); both \(P(C<X\mid X=3)\) and \(P(C<X\mid X=7)\) equal \(0.2500\).)
- In a validation sample, 12 of 20 high losses and 6 of 30 low losses are censored. Calculate the overall censoring rate and the rate difference between high and low losses. (Answer: Overall rate \(18/50=0.3600\); high rate \(0.6000\), low rate \(0.2000\), difference \(0.4000\).)
- A report states that 35% of records are censored but gives no censoring rule, loss-group counts, or information about how censoring times were set. Can the report establish noninformative censoring? State why. (Answer: No. The overall rate alone gives no relationship between \(X\) and \(C\), and does not establish their independence.)
- A hypothetical audit gives \(P(H)=0.5000\), \(P(E)=0.4000\), and \(P(H\cap E)=0.2000\), where \(H\) denotes high loss and \(E\) early threshold. Censoring occurs exactly for high losses with an early threshold. Is this joint distribution independent, and what is the censoring rate? (Answer: Yes, because \(P(H)P(E)=0.5000(0.4000)=0.2000=P(H\cap E)\); the censoring rate is \(0.2000\).)