A kernel regression model for panel count data with time-varying coefficients

Yang Wang, Zhangsheng Yu

Abstract

For the conditional mean function of panel count model with time-varying coefficients, we propose to use local kernel regression method for estimation. Partial log-likelihood with local polynomial is formed for estimation. Under some regularity conditions, strong uniform consistency rates are obtained for the local estimator. At target time point, we show that the local estimator converges in distribution to normal distribution. The baseline mean function estimator is also shown to be consistent. Simulation studies show that the time-varying coefficient estimator is close to the true value, the empirical coverage probabilities of the confidence interval is close to the nominal level. We also applied the proposed method to analyze a clinical study on childhood wheezing.

Introduction

In longitudinal studies, researchers often encounter "panel count data," where the exact timing of events is unknown. Instead, subjects are observed at discrete time points, and the number of events occurring between these points is recorded. Traditional statistical models for such data typically assume proportional means, implying that the effect of a covariate remains constant over the study period. However, in many biological and industrial contexts, the influence of risk factors is dynamic. For instance, the impact of a specific biomarker on disease progression may vary as a patient ages.

This paper addresses the limitations of constant-coefficient models by introducing a novel kernel regression approach. By allowing coefficients to vary over time, the model provides a more nuanced understanding of dynamic treatment effects. Unlike existing nonparametric methods like splines, which can be sensitive to knot placement, the proposed kernel-based method offers a robust framework with rigorous theoretical guarantees for consistency and asymptotic normality.

Methodology

The core of the proposed approach is local kernel regression. Rather than estimating a single global coefficient $\beta$, the method estimates a coefficient function $\beta(t)$ locally at each time point $t$.

Local Polynomial Approximation

The unknown coefficient function $\beta(u)$ is approximated in a small neighborhood of a target time $t$ using a Taylor expansion. This local polynomial approach allows the model to capture the curvature of the effect function accurately.

Kernel Weighting and Likelihood

A kernel function $K(\cdot)$ is employed to weight observations based on their proximity to the target time $t$. Observations closer to $t$ receive higher weights, effectively "localizing" the estimation process. The authors adapt the pseudo-likelihood approach, commonly used for Poisson processes, into a localized partial log-likelihood framework. By maximizing this local likelihood, they derive the estimator for the time-varying effect.

Theoretical Foundations

A significant contribution of this work is the establishment of the estimator's theoretical properties. The authors prove strong uniform consistency, ensuring that the estimate converges to the true function as the sample size grows. Furthermore, they demonstrate asymptotic normality, which facilitates the construction of reliable confidence intervals for the time-varying effects.

Simulation Results and Empirical Performance

The model's performance was evaluated through extensive simulation studies and a real-world application to childhood wheezing data.

Simulation Studies

Synthetic datasets were generated with known time-varying effects, such as square root curves and Beta density functions. The simulations demonstrated that the kernel estimator could accurately recover these underlying patterns.

Simulation results showing the accuracy of the time-varying coefficient estimator, standard errors, coverage probabilities, and the baseline mean function.

As shown in Figure 1, the estimated curves (a1, a2) closely track the true coefficient functions. The standard errors (a3, a4) remain low and stable across the time range, except near the boundaries. Crucially, the empirical coverage probabilities (a5, a6) are consistently near the nominal 95% level, validating the model's uncertainty estimates. The baseline mean function $\mu(t)$ (a7, a8) also shows strong consistency as the sample size increases from 300 to 500.

Application: Childhood Wheezing Study

The method was applied to a clinical study of 105 high-risk infants, tracking cumulative wheezing episodes over 33.5 months. The analysis focused on the effect of Interleukin-10 (IL-10), a cytokine involved in immune response. While a standard constant-coefficient model suggested a fixed effect of 0.428, the kernel model revealed that the risk associated with IL-10 is dynamic, peaking between 25 and 70 months. This insight provides a more detailed clinical picture of how immune markers influence respiratory health during early development.

Conclusion and Future Directions

The proposed kernel regression model offers a powerful tool for analyzing panel count data with dynamic effects. By moving beyond the assumption of constant coefficients, it enables researchers to identify critical windows of risk and treatment efficacy.

While the model is robust, its performance depends on the selection of the "bandwidth" parameter, which determines the degree of localization. Future research will focus on developing automated, data-driven techniques for bandwidth selection. Additionally, extending the model to handle departures from the Poisson process assumption, such as event clustering, remains an important area for further investigation.

Summary updated

Abstract

For the conditional mean function of panel count model with time-varying coefficients, we propose to use local kernel regression method for estimation.Partial log-likelihood with local polynomial is formed for estimation.Under some regularity conditions, strong uniform consistency rates are obtained for the local estimator.At target time point, we show that the local estimator converges in distribution to normal distribution.The baseline mean function estimator is also shown to be consistent.Simulation studies show that the time-varying coefficient estimator is close to the true value, the empirical coverage probabilities of the confidence interval is close to the nominal level.We also applied the proposed method to analyze a clinical study on childhood wheezing.
BibTeX
@inproceedings{wang2019kernel,
  author = {Yang Wang and Zhangsheng Yu},
  title = {{A kernel regression model for panel count data with time-varying coefficients}},
  booktitle = {Statistica},
  year = {2019},
  doi = {10.5705/ss.202019.0220},
}
Figure