Why Smooth a Censored Loss Distribution?
A right-censored record tells us that a loss exceeded a threshold, not its exact amount. For example, if a hypothetical policy has incurred a loss above a reporting threshold by the end of observation, its eventual amount is unknown. Treating that threshold as the actual loss puts an artificial point in the data and can distort a density estimate.
The Kaplan–Meier, or product-limit, estimate uses the observed failure times and the number of records still at risk at each time to estimate the loss distribution despite right censoring. Its estimate is a step function. Kernel smoothing replaces each estimated probability jump with a small, smooth density centered at that failure time. This makes the estimated density easier to interpret, but does not recover loss amounts that the data never reveal.
The standard interpretation assumes that censoring is independent of the loss of interest, so that being censored does not selectively hide larger or smaller losses after accounting for the study design. It also requires adequate follow-up: the loss distribution is estimable only over ranges where records remain under observation. With tied failures and censoring at the same time, use a consistent convention: failures at that time are counted among the at-risk records, and censorings at that time are removed after those failures are handled.
From Product-Limit Jumps to a Kernel Density
Choose a kernel density \(K(u)\), with cumulative distribution function \(\mathcal K(u)\), and a positive bandwidth \(h\). Its rescaled density is \(K_h(x-t)=K((x-t)/h)/h\). The kernel-smoothed density estimate is a weighted sum of kernels centered at observed loss times; the weights are the product-limit jumps, not equal weights assigned to all observed records.
Integrating this density up to \(x\) gives the corresponding smoothed distribution estimate:
These formulas are a kernel mixture: each estimated jump contributes a scaled kernel centered at its observed loss time. If the product-limit jumps sum to less than one, the smoothed density also integrates to less than one. That missing mass corresponds to estimated survival remaining beyond the observed failures; right-censored data alone do not say how that remaining mass is distributed at larger losses.
Worked Example: Smoothing Product-Limit Jumps
Worked Example: Smoothing Product-Limit Jumps
A hypothetical portfolio has eight loss records under observation at time 10. Two losses are observed at time 10 and one record is censored at that time. At time 20, one loss is observed among the five records still at risk. At time 30, one loss is observed among the four then at risk. The other three records are censored after time 30. Use a uniform kernel on \([-1,1]\) with bandwidth \(h=10\). Find the product-limit jump weights, the smoothed density at \(x=15\), and the total mass represented by the smoothed density.
Method. First calculate the product-limit survival and its probability decrement at each observed loss time. Then use those decrements as kernel weights. The censoring observation at time 10 is not counted as a loss.
At time 10, \(R_1=8\) and \(d_1=2\). Thus survival after time 10 is \(1-2/8=6/8=0.75\), and the jump weight is \(w_1=2/8=0.25\).
At time 20, \(R_2=5\) and \(d_2=1\). The survival immediately before time 20 is \(0.75\), so \(w_2=0.75(1/5)=0.15\). Survival after time 20 is \(0.75(4/5)=0.60\).
At time 30, \(R_3=4\) and \(d_3=1\). The survival immediately before time 30 is \(0.60\), so \(w_3=0.60(1/4)=0.15\). Survival after time 30 is \(0.60(3/4)=0.45\).
For the uniform kernel, \(K(u)=1/2\) when \(-1\le u\le1\), and \(K(u)=0\) otherwise. At \(x=15\), the standardized distances from the three event times are \((15-10)/10=0.5\), \((15-20)/10=-0.5\), and \((15-30)/10=-1.5\). The first two kernels contribute; the third does not. Therefore,
The total represented mass is \(0.25+0.15+0.15=0.55\). This agrees with \(1-\widehat S_{\rm KM}(30)=1-0.45=0.55\). The remaining estimated survival of \(0.45\) is not converted into an observed failure-time kernel. The density estimate is therefore not a complete density over all possible losses.
Check. The product-limit survival after the three event times is \(0.75(0.80)(0.75)=0.45\), and the jump weights sum to \(0.55\). The two calculations agree because the jumps account for the decrease in survival.
Why the Smoothing Formula Holds
The unsmoothed product-limit distribution assigns a probability jump \(w_j\) to each observed event time \(t_j\). In other words, its represented portion can be written as a sum of weighted point masses. Smoothing replaces each point mass by a kernel distribution with the same total weight and center \(t_j\).
The estimated probability assigned to event time \(t_j\) is \(w_j\). Thus the represented distribution is a mixture of point masses, with total mass \(\sum_j w_j\).
A kernel centered at \(t_j\) has density \(K_h(x-t_j)\), integrates to one, and is multiplied by the original weight \(w_j\). The resulting mixture density is \(\sum_j w_jK_h(x-t_j)\).
The integral of \(K_h(x-t_j)\) from negative infinity to \(x\) is \(\mathcal K((x-t_j)/h)\), by the change of variable \(u=(z-t_j)/h\). Summing the weighted integrals gives \(\widehat F_h(x)=\sum_jw_j\mathcal K((x-t_j)/h)\).
This argument also shows why smoothing preserves total represented mass: each kernel integrates to one, so \(\int \widehat f_h(x)\,dx=\sum_jw_j\). Smoothing changes the shape from jumps to a continuous curve; it does not change the amount of estimated probability represented by the observed jumps.
Worked Example: Evaluating the Smoothed Distribution
Worked Example: Evaluating the Smoothed Distribution
For a different hypothetical loss study, suppose the product-limit estimate has jumps \(0.20\), \(0.30\), and \(0.10\) at loss amounts \(40\), \(60\), and \(90\), respectively. Use a uniform kernel on \([-1,1]\) with bandwidth \(h=20\) to calculate the smoothed distribution and density estimates at \(x=55\).
For the uniform kernel, \(\mathcal K(z)=0\) for \(z\le-1\), \(\mathcal K(z)=(z+1)/2\) for \(-1<z<1\), and \(\mathcal K(z)=1\) for \(z\ge1\). The standardized distances are \((55-40)/20=0.75\), \((55-60)/20=-0.25\), and \((55-90)/20=-1.75\). The associated kernel CDF values are \(0.875\), \(0.375\), and \(0\). Therefore,
The uniform kernel density is \(1/(2h)=1/40=0.025\) inside its support. At \(x=55\), the kernels centered at 40 and 60 contribute, while the kernel centered at 90 does not. Hence,
The three jumps sum to \(0.60\), so the smoothed density integrates to \(0.60\), not one. If the product-limit survival after the last observed event is \(0.40\), this is consistent: that remaining survival is not represented by the three observed jumps. The common error to avoid is treating a partial product-limit estimate as if it necessarily specified the entire loss distribution.
Censoring and Tail Precision
Censoring affects tail estimates in two related ways. First, fewer records usually remain at risk at later loss times, so an observed failure can produce a relatively large product-limit decrement. Second, the data may stop providing event information while estimated survival is still positive. A kernel can spread observed tail jumps across nearby values, but it cannot establish where the unobserved remaining tail mass lies.
A standard pointwise uncertainty summary for the product-limit survival estimate is Greenwood’s approximation. At an observed time \(t\), it is
The formula concerns uncertainty in the product-limit survival estimate. It is not, by itself, a standard error formula for the kernel-smoothed density. As the risk sets become small, the Greenwood sum can grow, warning that the estimated tail is imprecise. The bandwidth also matters: a larger bandwidth makes the curve smoother but can blur meaningful local features, while a smaller bandwidth leaves more visible variation. Neither choice creates additional tail observations.
Worked Example: Checking Tail Uncertainty
In a hypothetical portfolio, two distinct event times have risk-set and failure counts \((R_1,d_1)=(15,3)\) and \((R_2,d_2)=(8,2)\). Calculate the product-limit survival estimate after the second time and its Greenwood standard error.
Method. Multiply the conditional survival factors to estimate survival, then substitute both risk sets into Greenwood’s formula.
After the first event time, survival is \(1-3/15=12/15=4/5\). After the second, it is \(1-2/8=6/8=3/4\) times the previous survival. Thus,
The two Greenwood terms are
Their sum is \(1/60+1/24=2/120+5/120=7/120\). Therefore the estimated variance is
The standard error is \(\sqrt{21/1000}\approx0.1449\), rounded to four decimal places. This is substantial relative to the survival estimate of \(0.60\), illustrating that a smooth-looking tail should not be mistaken for a precise one. The common mistake is to multiply the two survival factors incorrectly or to omit one of the risk-set contributions in Greenwood’s sum.
Common Exam Traps and Practical Interpretation
- Counting a censoring time as a failure. In the first worked example, the wrong calculation at time 10 would use \(d=3\) and report survival \(1-3/8=5/8\). Only two losses were observed, so the correct survival is \(1-2/8=6/8=0.75\). The censored record contributes information that the loss exceeded its censoring time, but it is not an observed loss at that time.
- Using equal weights for all records. Product-limit jumps depend on both the number at risk and earlier survival. A late observed loss does not automatically receive the same estimated probability as an early one.
- Reporting a proper density when mass remains unobserved. Check \(\sum_jw_j\). If it is below one, the estimated density from those jumps is a subdensity unless an additional tail model or explicit conditional interpretation is supplied.
- Interpreting smoothing as a cure for sparse follow-up. A kernel spreads probability around observed losses. It cannot determine the loss amounts of records censored before their losses were observed.
- Ignoring the scale of the density. Since \(K_h\) includes the factor \(1/h\), a density has units of inverse loss. Omitting that factor gives an incorrectly scaled estimate.
Check Your Understanding
Work each problem using the stated assumptions. For uniform kernels, use \(K(u)=1/2\) on \([-1,1]\) and zero outside.
- At one event time, \(R=12\) and \(d=3\). Find the product-limit survival immediately after the event and its probability jump. (Answer: survival \(0.75\); jump \(0.25\).)
- A product-limit estimate has jumps \(0.20\) at 10 and \(0.30\) at 20. With a uniform kernel and \(h=10\), find the smoothed density at \(x=15\). (Answer: \(0.025\) per loss unit.)
- At two event times the risk-set counts are \((R_1,d_1)=(20,4)\) and \((R_2,d_2)=(12,3)\). Find survival after the second time. (Answer: \(0.60\).)
- For risk-set counts \((15,3)\) and \((8,2)\), find the Greenwood standard error of survival after the second event time. (Answer: \(\sqrt{21/1000}\approx0.1449\).)
- A censored study has product-limit jumps totaling \(0.70\), with estimated survival \(0.30\) after the last observed event. What is the total integral of the density formed by smoothing those jumps, and what can be concluded about the remaining tail? (Answer: the integral is \(0.70\); the data do not locate or specify the distribution of the remaining \(0.30\) tail mass.)