Analysis of Longitudinal Data With Semiparametric Estimation of Covariance Function

Jianqing Fan, Tao Huang, Runze LiJournal of the American Statistical Association 2007 · DOI

TL;DR

This paper tackles the challenge of analyzing longitudinal data collected at irregular intervals by proposing a novel semiparametric approach to estimate the covariance function. This method improves the efficiency of regression coefficient estimation and trajectory prediction by combining flexible nonparametric variance estimation with a structured parametric correlation.

What Problem Does This Paper Address?

Analyzing longitudinal data, where measurements are taken repeatedly over time for the same individuals, is crucial for understanding individual trajectories and improving statistical model efficiency. A key component in these analyses is estimating the covariance function, which describes how measurements from the same individual are related over time. However, this becomes significantly more complex when data points are not collected at fixed, regular intervals, or when these intervals vary across individuals. Existing methods often struggle with such irregular data, leading to less efficient or even inconsistent estimates. This paper aims to bridge this gap by developing a robust method for estimating the covariance function in the presence of irregular and subject-specific time points.

Methods

The core of the proposed methodology lies in a semiparametric model for the covariance function. This model separates the covariance into two parts:

  • Nonparametric Variance Function: The variance of the error term at any given time point, $\sigma^2(t)$, is estimated using a flexible kernel smoothing approach. This allows the variance to change over time without imposing a rigid parametric form.
  • Parametric Correlation Structure: The correlation between error terms at different time points, $\text{corr}(\epsilon(s), \epsilon(t))$, is modeled using a parametric function $p(s, t, \theta)$. This imposes a structure on the correlation, which is crucial for ensuring positive definiteness of the estimated covariance matrix and improving efficiency.

Two methods are proposed for estimating the parameters ($\theta$) of this correlation structure:

  1. Quasi-Likelihood (QL) Estimator: This method maximizes a likelihood-like function based on the assumed parametric correlation structure.
  2. Minimum Generalized Variance (MGV) Estimator: This approach aims to directly minimize the generalized variance of the regression coefficient estimator, effectively seeking to minimize the volume of the confidence ellipsoid for $\beta$.

For estimating the regression coefficients ($\alpha(t)$ and $\beta$), the paper introduces a profile weighted least squares approach. This method iteratively refines estimates by first estimating the nonparametric components and then using these estimates to improve the estimation of the parametric components.

Datasets

The paper uses a subset of data from the Multi-Center AIDS Cohort (MACS) study for illustration in Section 5.3. This dataset contains HIV status and related covariates for 283 homosexual men from 1984 to 1991. The data is characterized by irregular measurement times and missed visits, making it a suitable test case for the proposed methods.

Key Insights and Ideas

  • Balancing Flexibility and Structure: The semiparametric approach is a key insight, allowing for the flexibility of nonparametric variance estimation while retaining the desirable properties (like positive definiteness and efficiency gains) of a structured parametric correlation.
  • Irregular Data is a Major Hurdle: The paper highlights that irregular and subject-specific time points are a significant challenge in longitudinal data analysis, motivating the need for robust estimation techniques.
  • MGV vs. QL: The Minimum Generalized Variance (MGV) method is shown to be more stable and robust than the Quasi-Likelihood (QL) method, particularly in cases of low correlation, suggesting it's a more reliable choice when the exact correlation structure might be uncertain.
  • Efficiency Gains: Incorporating the correlation structure, even if not perfectly specified, can lead to substantial efficiency gains in estimating regression coefficients compared to assuming independence.
  • Model Flexibility: The proposed varying-coefficient partially linear model offers a more flexible framework than traditional linear models for longitudinal data, allowing for time-varying effects.

Key Results

  • Asymptotic Normality: The paper establishes the asymptotic normality of the proposed estimators for both regression coefficients and the variance function, providing a theoretical foundation for their use.
  • Simulation Performance: Monte Carlo simulations demonstrate that the proposed semiparametric methods perform well. The QL and MGV estimators for $\beta$ are significantly more efficient than using a working independence correlation structure, with relative efficiencies ranging from 1.3 to 3 for different correlation strengths. The MGV method showed better stability.
  • Real Data Application: The analysis of the MACS data successfully illustrates the application of the proposed methodology, providing insights into the trends of CD4 cell percentage and the time-varying impact of PreCD4.
  • Standard Error Accuracy: The proposed formula for estimating standard errors (Equation 3.7) was found to be accurate for both correctly specified and misspecified correlation structures.

Limitations and Future Work

  • Robustness of QL: The paper notes that the QL criterion might not be as robust as MGV, especially when the correlation structure is misspecified, and the optimization algorithm may not always converge.
  • Nonparametric Correlation: The paper focuses on parametric correlation structures. While a nonparametric estimate for the correlation function is presented (Equation 6.1), its properties (like positive definiteness) are not guaranteed, and it might be less useful for small sample sizes or unbalanced data.
  • Sampling Properties of $\theta$: The paper does not fully discuss the sampling properties of the estimators for $\theta$ derived by QL and MGV. Investigating these properties, especially when the correlation function is misspecified, is suggested as an area for future research.
  • Robust Estimation: While the paper mentions robust methods for semiparametric regression, it doesn't explicitly incorporate them into the proposed covariance estimation framework. Further research into robust estimation procedures for $\alpha(t)$ and $\beta$ in the presence of outliers is suggested.
  • Alternative Estimation Strategies: The paper briefly mentions decorrelating data first and then applying profile least squares as an alternative strategy for estimating $\beta$, suggesting further research and comparison would be beneficial.

Summary updated

Abstract

Improving efficiency for regression coefficients and predicting trajectories of individuals are two important aspects in analysis of longitudinal data. Both involve estimation of the covariance function. Yet, challenges arise in estimating the covariance function of longitudinal data collected at irregular time points. A class of semiparametric models for the covariance function is proposed by imposing a parametric correlation structure while allowing a nonparametric variance function. A kernel estimator is developed for the estimation of the nonparametric variance function. Two methods, a quasi-likelihood approach and a minimum generalized variance method, are proposed for estimating parameters in the correlation structure. We introduce a semiparametric varying coefficient partially linear model for longitudinal data and propose an estimation procedure for model coefficients by using a profile weighted least squares approach. Sampling properties of the proposed estimation procedures are studied and asymptotic normality of the resulting estimators is established. Finite sample performance of the proposed procedures is assessed by Monte Carlo simulation studies. The proposed methodology is illustrated by an analysis of a real data example.
BibTeX
@article{fan2007analysis,
  author = {Jianqing Fan and Tao Huang and Runze Li},
  title = {{Analysis of Longitudinal Data With Semiparametric Estimation of Covariance Function}},
  journal = {Journal of the American Statistical Association},
  year = {2007},
  doi = {10.1198/016214507000000095},
}
Figure