Skip to contents

Estimation and inference when a longitudinal covariate is observed sparsely and intermittently, at times that do not line up with the times at which the outcome is measured. For a survival outcome, fits the transformed hazards family – proportional hazards, additive hazards, and the Box-Cox transformations between them – by the Sieve Maximum Kernel-weighted Log-likelihood Estimator (SMKLE) of Sun, Sun, Zhao and Cao (2025) doi:10.1080/01621459.2025.2476781 , with specialised kernel estimating-equation alternatives for the proportional and additive cases. For a longitudinal outcome recorded on its own time grid, fits generalised linear models by the kernel-weighted estimating equations of Cao, Zeng and Fine (2015) doi:10.1111/rssb.12086 , with either time-invariant or time-dependent coefficients. Half and full kernels, data simulation, and bandwidth selection by cross-validation are available throughout, and the numerical routines use 'Rcpp', 'RcppArmadillo' and the 'nloptr' C API.

Details

A longitudinal covariate is rarely measured at the time you need it. skmle handles that mismatch directly, by weighting each observation according to how far its measurement time sits from the time being modelled, rather than carrying a value forward or smoothing the covariate and substituting it. Two outcome types are covered.

Which function do I need?

Answer two questions.

First, what is the outcome?

A time to an event

(death, relapse, failure), possibly censored – use the survival estimators. Your data is one long table with a row per covariate measurement.

A repeatedly measured quantity

(a score, a lab value) recorded on its own schedule – use the asynchronous estimators. Your data is two tables, one per process.

Then pick within that group.

OutcomeSituationFunction
SurvivalStart here; Cox modelkee_cox()
SurvivalAdditive hazards insteadkee_additive()
SurvivalWant the baseline hazard, or a model between the twoskmle()
SurvivalChoose the bandwidth properlyskmle_cv()
LongitudinalStart here; one constant effectkee_async()
LongitudinalChoose the bandwidth properlykee_async_cv()
LongitudinalThe effect may change over timekee_async_td()

Every fitting function will pick a bandwidth for you if you do not supply one, and will say in a message what it chose. That is enough to get a first answer; the _cv() functions choose it from the data, which is what to report.

Survival outcomes

skmle() fits the transformed hazards family by the Sieve Maximum Kernel-weighted Log-likelihood Estimator of Sun, Sun, Zhao and Cao (2025). The Box-Cox parameter s indexes the family: s = 0 is proportional hazards, s = 1 is additive hazards, and other values interpolate. kee_cox() and kee_additive() are faster specialised estimating equations for the two named cases, and skmle_cv() selects the bandwidth by subject-level cross-validation. Simulate with sim_skmle_data().

Asynchronous longitudinal outcomes

When the outcome is itself a sparsely observed longitudinal process, recorded on a time grid that does not line up with the covariate's, kee_async() and kee_async_td() fit generalised linear models by the kernel-weighted estimating equations of Cao, Zeng and Fine (2015), with time-invariant coefficients and with a coefficient curve \(\beta(t)\) respectively. Simulate with sim_async_data().

Half and full kernels

Every estimator takes one_sided. A half kernel admits only measurements strictly before the time being modelled, which is the risk-set restriction of a hazard model and the causal reading of a covariate path; a full kernel smooths from both sides. The survival estimators default to the half kernel, as published; the asynchronous ones default to the full kernel, as published.

References

Sun, D., Sun, Z., Zhao, X. and Cao, H. (2025). Kernel meets sieve: transformed hazards models with sparse longitudinal covariates. Journal of the American Statistical Association 120, 2580-2591.

Cao, H., Zeng, D. and Fine, J. P. (2015). Regression analysis of sparse asynchronous longitudinal data. Journal of the Royal Statistical Society, Series B 77, 755-776.

Author

Maintainer: Dayu Sun dayusun@iu.edu [copyright holder]

Authors: