A non-parametric estimator for excess recurrent events
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "A non-parametric estimator for excess recurrent events".
Jane: Here's the gist of the paper. Measuring disease burden is an important part of both public health research and health economics.
Tom: This paper extends the notion of relative survival to recurrent events by developing a novel nonparametric estimator.
Jane: The estimator combines data from some cohort with aggregated population level data to estimate the number of excess recurrent events.
Tom: Hospitalisations and other recurrent events occur frequently in epidemiological data.
Paper summary: Tom: Looking at the full title, "A non-parametric estimator for excess recurrent events," it tells you immediately that this work is focused on finding a statistical tool that doesn't rely on assuming a specific mathematical shape for how these events happen Tom. It's about building a method that works without having to pick a fixed model beforehand Jane.
Jane: And the authors, Hedberg and Hössjer, are using this estimator to estimate the number of excess recurrent events by combining data from some cohort with aggregated population level data Lu. The implication is that this gives us a way to quantify how much more often certain events occur in a diseased group compared to what we'd expect from the general public Jane.
Lu: What this means practically is that instead of just counting hospitalizations, you can estimate the actual excess burden on the healthcare system by subtracting what would have happened without those specific disease factors Lu. It helps researchers bypass issues where recorded reasons for events don't accurately capture the total disease impact Tom.
Meng: From a production standpoint, if this method is robust—and they show consistency for the covariance function—it means we can get a more reliable number on how many extra cases you expect to see in a population with that condition Meng. It moves us away from potentially misleading counts based only on labeled reasons Jane.
Lalam: And the paper concludes by showing that this estimator is useful for quantifying the burden of illness across different diseases, even suggesting stratification might be better for individual patient decisions Lalam. It's a general tool for measuring excess events in many contexts Tom.
Jane: So, in simple terms, this paper gives us a statistically sound way to measure the extra events that happen because of a disease compared to normal population trends, even when the reasons are messy Jane. It's about getting a better handle on the real impact of recurrent illness Tom.
Conclusion: Tom: So, we've been looking at this paper titled "A non-parametric estimator for excess recurrent events." Basically, they’ve developed a way to measure how many extra events you see because of a disease compared to what happens in the general population without that disease.
Jane: It sounds like it’s trying to solve the problem where we just count things, but we don't know if those counts truly reflect the total burden of the illness.
Tom: Exactly. And they use this estimator by combining data from a specific group with broader population numbers to figure out that excess number.
Lu: The core idea is decomposing those individual event rates into an excess rate and a general population rate, which lets them estimate the excess part without needing a fixed model for how those events follow over time.
Meng: I'm curious about the practical side—how reliable is this estimate when you actually try to deploy it on real data at scale?
Jane: That’s a good question, Meng. The authors did some simulations, and they showed that the estimator and its error bounds hold up pretty well in tests mimicking real registry scenarios.
Tom: Yeah, I saw those results where the expected number of excess events estimator behaved as expected even with sample sizes of five hundred and two thousand. It seems consistent enough for this kind of work.
Lu: And they proved that the mathematical process they are using converges to a zero mean Gaussian process, which gives them a strong theoretical foundation for how to estimate the uncertainty in their results.
Jane: So, while it’s not a magic fix for everything, it provides a solid framework for quantifying that extra burden on patients and the healthcare system.
Tom: It definitely seems like they're pointing toward using this kind of statistical tool when you need to compare recurrent events against expected population rates for diseases with many recurring issues.
Jane: And this leads us to thinking about what kind of diseases might benefit most from this kind of measurement, especially given the focus on high mortality scenarios.
Meng: It’s interesting that they mention it's relevant for cancers or heart failure—diseases where the recurrence and death rates are both quite high.
Lu: The paper suggests a way to isolate excess recurrent events and excess mortality separately, which is a big step because often those two things get mixed up in standard survival analysis.
Tom: It’s less about finding a single perfect answer and more about building this tool that helps us measure the discrepancy between what we see in a patient group versus the baseline population expectation.
Department of Medical Epidemiology and Biostatistics, Karolinska Institutet · Department of Mathematics, Stockholm University · Department of Molecular Medicine and Surgery, Karolinska Institutet
stat.ME
Submitted: 2026-09-22
Updated: 2026-11-09
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 82/100
The gist: Measuring disease burden is an important part of both public health research and health economics.
Key concepts
- Relative Survival
- A survival analysis method that adjusts observed survival rates in a specific group (cohort) using data from the entire population. It helps account for differences in mortality rates between groups, allowing researchers to compare outcomes more fairly.
- Excess Recurrent Events
- This refers to the number of recurrent events (like hospitalizations) that happen more frequently than expected in a population without the specific disease. The estimator aims to quantify this 'excess' burden.
- Non-parametric Estimator
- A statistical method that does not assume a specific mathematical distribution for the data, unlike parametric methods. This estimator uses product limit estimators to estimate survival and event rates directly from the data, making it flexible for various disease patterns.
Terminology
Summary
Measuring disease burden is an important part of both public health research and health economics. p. 0 A non-parametric estimator for excess recurrent
events
Jonatan Hedberg1
, Ola H¨ossjer2
, Caroline Nordenwall3 Therese
M-L Andersson1
, Elisavet Syriopoulou1
1Department of Medical Epidemiology and Biostatistics, Karolinska Institutet,
Stockholm, Sweden
2Department of Mathematics, Stockholm University, Stockholm, Sweden
3Department of Molecular Medicine and Surgery, Karolinska Institutet, Stockholm,
Sweden
September 23, 2026
Abstract
Measuring disease burden is an important part of both public health research and health economics. Some measures of disease burden, such as
hospitalisations, are made up of recurrent events. Assessing what events
are related to a particular disease is however non-trivial. We extend the
notion of relative survival to recurrent events by developing a novel nonparametric estimator. We extend the
notion of relative survival to recurrent events by developing a novel nonparametric estimator. The estimator combines data from some cohort
with aggregated population level data to estimate the number of excess
recurrent events.Where the paper says it
Where the paper says it
Where the paper says it
Introduction and Motivation
Hospitalisations and other recurrent events occur frequently in epidemiological data. The need for methods that can bypass the problem of recorded reasons not accurately capturing disease burden is highlighted. p. 1 Only measuring the hospitalisations that are
labelled as cancer-related would then result in an underestimation of the disease
burden. Therefore, there is a need for methods and concepts that can bypass
this problem. Relative survival methods deal with this problem in the survival analysis
setting. For these methods population level mortality data (usually from life
tables) are used to adjust the observed survival in the cohort currently being
analysed. To our knowledge, there is no literature at the moment containing methods for how to handle relative recurrent events. We therefore propose a nonparametric estimator that allows for the estimation of the excess number of
recurrent events in relation to what would be expected in a population without
the disease in question.Where the paper says it
Where the paper says it
Where the paper says it
The Non-Parametric Estimator
The estimator is formulated based on an additive excess rate model where individual recurrent events rates are decomposed into an excess rate and a general population rate. The expected number of excess recurrent events up to time t is given by the formula µE(t) = Z t 0 dΛR E(u)S(u) (1). p. 2 Following the same thinking as in Ghosh & Lin (2000), we are interested in
knowing how many recurrent events we expect to occur while individuals are
still alive. For this reason we want to estimate the expected number
µE(t) = Z t
0
dΛ
R
E(u)S(u) (1)
of excess recurrent events up to time t. We propose estimating µE(t) by plugging in marginal non-parametric estimates of S(t) and dΛ
R
E(t) into (1). These estimates are given by the product
limit estimator
Sˆ(t) = P
[0,t]
1 −
Pn
i=1 dN D
i
(s)
Pn
i=1 Yi(s)
= P
[0,t]
1 − dΛˆD(s)
(2)
and
dΛˆR
E(t) = Xn
i=1
dN R
i
(t) − Yi(t)dΛ
R
P
(t; Xi)
Pn
j=1 Yj (t)
, (3)
respectively. This is the marginal number of excess
recurrent events that a randomly sampled individual with the disease suffers
during [0, t] as long as they are alive (whereafter they can have no such events). We propose estimating µE(t) by plugging in marginal non-parametric estimates of S(t) and dΛ
R
E(t) into (1).Where the paper says it
Where the paper says it
Where the paper says it
Theoretical Framework
The theoretical contribution involves three theorems detailing the weak convergence of the estimator to a zero mean Gaussian process and how to estimate its asymptotic covariance function. Theorem 1 shows that the right continuous version of a specific process converges weakly as n → ∞ to a multivariate zero mean Gaussian process. p. 3 The right continuous version of the
process {
√
n(N R(t) − E[N R(t)], N D(t) − E[N D(t)], N R
P
(t) − E[N R
P
(t)], Y (t) −
E[Y (t)]); 0 ≤ t ≤ τ} then converges weakly as n → ∞ to a multivariate zeromean Gaussian process {U(t); 0 ≤ t ≤ τ} on the metric space (D
4
[0, τ ], ||·||∞) of
four-dimensional c`adl`ag functions on [0, τ ], equipped with the supremum norm. The second theorem contains the main result of the paper. Then as n → ∞ the process
{
√
n(ˆµE(t) − µE(t)); 0 ≤ t ≤ τ} converges weakly to a zero mean Gaussian process W = {W(t); 0 ≤ t ≤ τ}, with covariance function v(s, t) = Cov[W(s), W(t)],
on (D[0, τ ], ||. Our last theorem follows more or less directly from the proof of theorem 2
together with some additional arguments regarding consistency. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞.Where the paper says it
Where the paper says it
Where the paper says it
Simulation and Application
A simulation study was conducted to evaluate the finite sample properties of the estimator using data mimicking registry scenarios. The simulations, performed for sample sizes n = 500 and n = 2000, showed that the expected number of excess recurrent events estimator ˆµE(t) and its standard error perform well in finite samples. p. 6 We see that the
expected number of excess recurrent events estimator ˆµE(t), and its standard
error ˆσ(t, t), perform well in finite samples. The actual coverage of the confidence
intervals I(t) is generally close to their nominal coverage for both the smaller
and the larger sample size, although some tendencies of overestimation of the
standard deviation can be seen in a few scenarios. We also evaluate the finite sample properties of the estimator through simulations and provide a practical example using data from
Swedish patients with rectal cancer. Keywords: Excess recurrent events; Non-parametric estimator; Relative
survival
1 Introduction
Hospitalisations and other recurrent events occur frequently in epidemiological
data. At the end of study the estimated number
of excess hospitalisations was 4. 87, with a 95% confidence interval (4. 87, with a 95% confidence interval (4. This indicates a substantial burden on both patients and the healthcare system.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Discussion and Interpretation
The estimator is based on a marginal interpretation, averaging over both disease related and non-disease related patient characteristics to provide an estimate of the total burden of disease. The authors note that the estimator does not produce a causal estimate because overall survival is part of the estimand. p. 9 Another aspect related to the interpretation our estimator is that it does not
produce a causal estimate in the sense of what would happen should one be able
to isolate the impact of the disease. The reason being that overall survival is
part of the estimand. In particular, the estimator is highly relevant for
those diseases that are characterised both by a large number of some recurrent
events and high mortality. Examples of such diseases include certain cancer
types, as illustrated here, heart failure, and chronic pulmonary disease. This is a conscious choice on our behalf, since we believe
that there is greater benefit in being able to quantifying something that is fully
interpretable in the real world - the excess number of events in comparison to the
general population taking death into account. It is nonetheless certainly possible
to replace the product limit estimator (2) in (4) with a suitable estimator of
relative survival.Where the paper says it
Where the paper says it
Where the paper says it
Supplementary Results
The proof of theorem 3 shows that the covariance function estimator vˆ(s, t) = nσˆ2(s, t) is a consistent estimator of the asymptotic covariance function v(s, t). The convergence is uniform for all (s, t) in the sense that max 0≤s,t≤τ vˆ(s, t) − v(s, t) P→ 0 as n → ∞. p. 4 Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. The
convergence is uniform for all (s, t) in the sense that
max
0≤s,t≤τ
|vˆ(s, t) − v(s, t)|
P
→ 0 (8)
as n → ∞. We see that the
expected number of excess recurrent events estimator ˆµE(t), and its standard
error ˆσ(t, t), perform well in finite samples. The actual coverage of the confidence
intervals I(t) is generally close to their nominal coverage for both the smaller
and the larger sample size, although some tendencies of overestimation of the
standard deviation can be seen in a few scenarios. The percentage of patients
with stage 3 was 24. The estimated number ˆµE(t) of excess hospitalisations is displayed in figure
1 as a function of t (given in days). In figure 1 we see an accumulation of excess events that decreases in speed as
time passes. The decrease is of no surprise, due to the high mortality of this
patient group, with 90% dying before censoring. some sort of recurrent events
deficit. The multiplication with the estimated survival curve {Sˆ(u); 0 ≤ u ≤ τ}
in the integrand of ˆµE(t) does not change the interpretation of the estimated
relative recurrences in (4), since individuals after their death will always have
zero expected number of recurrent events. This might be
seen as a downside of the estimator but we would argue that it is not. The
estimator is simply comparing the number of recurrent events individuals have
with what we expect based on the general population. Treatment related admissions
are especially likely during the first year since diagnosis. We can therefore
conclude that there is a substantial burden of disease in this patient group and
that earlier detection would be of immense value both to society and patients. size in each stratum is large enough. Regression modelling might however be
preferable. It is nonetheless certainly possible
to replace the product limit estimator (2) in (4) with a suitable estimator of
relative survival. This could under certain circumstances allow for the isolation
of both excess recurrent events and excess mortality. some sort of recurrent events
deficit. The multiplication with the estimated survival curve {Sˆ(u); 0 ≤ u ≤ τ}
in the integrand of ˆµE(t) does not change the interpretation of the estimated
relative recurrences in (4), since individuals after their death will always have
zero expected number of recurrent events. Lastly, we
illustrated how the estimator might be used in Swedish data on patients with
rectal cancer. It is our belief that this estimator will prove useful in quantifying the burden
of illness of different diseases. It is perhaps of lesser value to individual patients and
for decision making by physicians, since we would then have at least partial information on the characteristics of the particular patient. This can be partially
remedied through stratification, e.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
References
Andersen, P. K., & Væth, M. p. 1 There is much literature on this subject. Several non-parametric
estimators for relative survival exist (Andersen & Væth, 1989; Pohar Perme
et al. Several non-parametric
estimators for relative survival exist (Andersen & Væth, 1989; Pohar Perme
et al. , 2012), and the concept has been extended to regression models both
semi-parametric (Sasieni, 1996) and parametric (Nelson et al.Where the paper says it
Where the paper says it
Eddelbuettel, D. p. 5 Anticipating application of the method using registry data where individuals
are enrolled into a study at the date of diagnosis, we created simulations that
mimic such a scenario. Computations were done in C++ and R through the
Rcpp interface (Eddelbuettel et al. Assume conditions 1-3 above. The right continuous version of the
process {
√
n(N R(t) − E[N R(t)], N D(t) − E[N D(t)], N R
P
(t) − E[N R
P
(t)], Y (t) −
E[Y (t)]); 0 ≤ t ≤ τ} then converges weakly as n → ∞ to a multivariate zeromean Gaussian process {U(t); 0 ≤ t ≤ τ} on the metric space (D
4
[0, τ ], ||·||∞) of
four-dimensional c`adl`ag functions on [0, τ ], equipped with the supremum norm.Where the paper says it
Where the paper says it
Ghosh, D., & Lin, D. p. 2 The hazard rate at time t of a randomly chosen individual from
population is given by dΛ
D(t) = E[Yi(t)dΛ
D(t; Xi
, Zi)]/[C(t)S(t)]. Following the same thinking as in Ghosh & Lin (2000), we are interested in
knowing how many recurrent events we expect to occur while individuals are
still alive. Assessing what events
are related to a particular disease is however non-trivial. We extend the
notion of relative survival to recurrent events by developing a novel nonparametric estimator.Where the paper says it
Where the paper says it
Gill, R. D., & Johansen, S. (1990). A survey of product-integration with a view toward application in survival analysis. p. 1 Therefore, there is a need for methods and concepts that can bypass
this problem. Relative survival methods deal with this problem in the survival analysis
setting.Where the paper says it
Lindqvist, B. p. 6 If Ni(tk−1, tk) > 0 generate the sum of the event times from a truncated
Gamma distribution. Generate exponential times based on the sum of event times as described
in Lindqvist & Taraldsen Lindqvist and Taraldsen, 2007. If Ni(tk−1, tk) > 0 generate the sum of the event times from a truncated
Gamma distribution. Generate exponential times based on the sum of event times as described
in Lindqvist & Taraldsen Lindqvist and Taraldsen, 2007. The
scenarios were investigated for the sample sizes n = 500 and n = 2000 and each
scenario was repeated 5000 times per sample size. This should give a Monte
Carlo error of around 0.Where the paper says it
Where the paper says it
Where the paper says it
Nelson, C. P., Lambert, P. C., Squire, I. B., & Jones, D. R. (2007). Flexible parametric models for relative survival, with application in coronary heart disease. p. 1 Several non-parametric
estimators for relative survival exist (Andersen & Væth, 1989; Pohar Perme
et al. , 2012), and the concept has been extended to regression models both
semi-parametric (Sasieni, 1996) and parametric (Nelson et al.Where the paper says it
Pohar Perme, M. p. 1 There is much literature on this subject. Several non-parametric
estimators for relative survival exist (Andersen & Væth, 1989; Pohar Perme
et al. 5 Discussion
We have extended the concept of relative survival to the field of recurrent events
by presenting a non-parametric estimator for the excess number of recurrent
events. This estimator allows for the estimation of the societal burden e.Where the paper says it
Where the paper says it
Sasieni, P. p. 1 Several non-parametric
estimators for relative survival exist (Andersen & Væth, 1989; Pohar Perme
et al. , 2012), and the concept has been extended to regression models both
semi-parametric (Sasieni, 1996) and parametric (Nelson et al. A non-parametric estimator for excess recurrent
events
Jonatan Hedberg1
, Ola H¨ossjer2
, Caroline Nordenwall3 Therese
M-L Andersson1
, Elisavet Syriopoulou1
1Department of Medical Epidemiology and Biostatistics, Karolinska Institutet,
Stockholm, Sweden
2Department of Mathematics, Stockholm University, Stockholm, Sweden
3Department of Molecular Medicine and Surgery, Karolinska Institutet, Stockholm,
Sweden
September 23, 2026
Abstract
Measuring disease burden is an important part of both public health research and health economics. Some measures of disease burden, such as
hospitalisations, are made up of recurrent events.Where the paper says it
Where the paper says it
van der Vaart, A. p. 3 As a side note the convergence in probability in (5) can be
interpreted as ordinary probabilities and not in terms of outer probabilities (cf. van der Vaart and Wellner, 2023) since
µˆE(t) and µE(t) are both measurable functions of t on the compact set
t ∈ [0, τ ]. As a side note the convergence in probability in (5) can be
interpreted as ordinary probabilities and not in terms of outer probabilities (cf. van der Vaart and Wellner, 2023) since
µˆE(t) and µE(t) are both measurable functions of t on the compact set
t ∈ [0, τ ]. These four processes are the building blocks of ˆµE = {µˆE(t); 0 ≤ t ≤ τ} in (4). To ensure weak
convergence of these four processes and ˆµE, we make the following assumptions:
1.Where the paper says it
Where the paper says it
Where the paper says it
Weibull, C. p. 7 95
admissions can be found. For more information on CRCBaSe and its data
sources see Weibull et al. 4 Application
We illustrate the usage of the relative recurrent events estimator using data
from CRCBaSe, a Swedish registry containing data on patients with colon or
rectal cancer. This database contains data on around 77,000 patients diagnosed
with colon or rectal cancer taken from SCRCR, the national quality registry
for colorectal cancer.Where the paper says it
Where the paper says it
Appendix
A.1 Proof of theorem 1
Proof. It is convenient to introduce the process Un = Un(t), with Un(t) = √n(N R(t) − E[N R(t)], √n(N D(t) − E[N D(t)]), = √n(N R P (t) − E[N R P (t)], √n[Y (t) − E[Y (t)]])=: Un1(t), Un2(t), Un3(t), Un4(t). p. 0 The estimator combines data from some cohort
with aggregated population level data to estimate the number of excess
recurrent events. Using empirical process theory, we show that the estimator converges weakly to a mean zero Gaussian process under mild
regularity conditions, and provide a consistent estimator for the covariance function. Assume conditions 1-3 above. The right continuous version of the
process {
√
n(N R(t) − E[N R(t)], N D(t) − E[N D(t)], N R
P
(t) − E[N R
P
(t)], Y (t) −
E[Y (t)]); 0 ≤ t ≤ τ} then converges weakly as n → ∞ to a multivariate zeromean Gaussian process {U(t); 0 ≤ t ≤ τ} on the metric space (D
4
[0, τ ], ||·||∞) of
four-dimensional c`adl`ag functions on [0, τ ], equipped with the supremum norm.Where the paper says it
Where the paper says it
A.2 Proof of theorem 2
Proof. Letting N R, N R P, N D and Y be represented by A, B, G and D the estimator µˆE = (µˆE(t); 0 ≤ t ≤ τ) consists of the following maps: A → A − B, G. p. 3 These four processes are the building blocks of ˆµE = {µˆE(t); 0 ≤ t ≤ τ} in (4). To ensure weak
convergence of these four processes and ˆµE, we make the following assumptions:
1. Assume conditions 1-4 above. Then as n → ∞ the process
{
√
n(ˆµE(t) − µE(t)); 0 ≤ t ≤ τ} converges weakly to a zero mean Gaussian process W = {W(t); 0 ≤ t ≤ τ}, with covariance function v(s, t) = Cov[W(s), W(t)],
on (D[0, τ ], ||.Where the paper says it
Where the paper says it
A.3 Proof of theorem 3
Proof. The corresponding estimator of the covariance function vˆ(s, t) = nσˆ2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Sˆ(u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u). p. 4 This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞. This gives rise to an estimate
σˆ
2
(s, t) = 1
n2
Xn
i=1
Φˆ
i(s)Φˆ
i(t) (6)
of the covariance function, with
Φˆ
i(t) = Z t
0
Sˆ(u)
dN R
i
(u) − Yi(u)d(ΛR
P
(u; Xi) + ΛˆR
E(u))
Y (u)
− µˆE(t)
Z t
0
dN D
i
(u) − Yi(u)dΛˆD(u))
Y (u)
(7)
+
Z t
0
µˆE(u)
dN D
i
(u) − Yi(u)dΛˆD(u)
Y (u)
. Moreover, vˆ(s, t) = nσˆ
2
(s, t) is a consistent estimator of the asymptotic covariance function v(s, t) of the Gaussian process W in theorem 2 as n → ∞.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The corresponding estimator of the covariance function vˆ(s, t) = nσ 2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Ŝ (u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u). The corresponding estimator of the covariance function vˆ(s, t) = nσ 2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Ŝ (u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u). The corresponding estimator of the covariance function vˆ(s, t) = nσ 2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Ŝ (u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u).
The corresponding estimator of the covariance function vˆ(s, t) = nσ 2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Ŝ (u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u). The corresponding estimator of the covariance function vˆ(s, t) = nσ 2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Ŝ (u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u).
The corresponding estimator of the covariance function vˆ(s, t) = nσ 2(s, t) is 1/n Xn i=1 Φˆi(s)Φˆi(t), where Φˆi(t) = Z t 0 Ŝ (u) dN R i (u − Yi (u)-d[ΛR P (u; Xi) + ΛˆR E (u)] Y (u)— µ̂E(t) Z t 0 dN D i (u − dYi(t)ΛˆD(t)) Y (u) + Z t 0 µ̂E(u) dN D i (u − Yi (u)-dΛˆD(t)) Y (u).
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved disease burden estimation for recurrent events by developing a novel nonparametric estimator that combines cohort data with aggregated population level data to estimate excess recurrent events. This allows for quantification of disease burden without needing a control group, extending relative survival concepts to recurrent events.
-
Implementation of a weak convergence proof using empirical process theory to show the estimator converges weakly to a mean zero Gaussian process under mild regularity conditions. This provides theoretical grounding for the estimator's asymptotic behavior and consistency in functional terms.
-
Consistent estimation of the covariance function for excess recurrent events by providing an estimate based on finite sample properties, specifically using the formula sigma two (s, t) = 1/n2 Xn i=1 E[Φi(s)Φi(t)] and its estimator vˆ(s, t) = nσˆ2 (s, t). This enables the construction of confidence intervals for disease burden estimates.
-
Application of the estimator to real-world registry data to estimate the excess number of hospitalizations caused by a disease in patients with stage 3 or 4 rectal cancer. The system can now produce estimated numbers and associated confidence intervals for specific time points, such as at the end of study.
-
Ability to interpret negative results in terms of a
recurrent events deficit,
where the comparison is against what would be expected based on the general population taking death into account, rather than relying solely on log transforms.
Abstract
Measuring disease burden is an important part of both public health research and health economics. Some measures of disease burden, such as hospitalisations, are made up of recurrent events. Assessing what events are related to a particular disease is however non-trivial. We extend the notion of relative survival to recurrent events by developing a novel non-parametric estimator. The estimator combines data from some cohort with aggregated population level data to estimate the number of excess recurrent events. Using empirical process theory, we show that the estimator converges weakly to a mean zero Gaussian process under mild regularity conditions, and provide a consistent estimator for the covariance function. We also evaluate the finite sample properties of the estimator through simulations and provide a practical example using data from Swedish patients with rectal cancer.
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes