Local polynomial regression for pooled response data

Dewei Wang, Xichen Mou, Xiang Li, Xianzheng Huang

TL;DR

This paper introduces new statistical methods for analyzing data where individual measurements are grouped into "pools" before being measured. These methods, called local polynomial regression for pooled data, allow researchers to estimate relationships between variables even when only the average response within a pool is known, which is common in fields like epidemiology and environmental science due to cost or sample preservation.

What Problem Does This Paper Address?

In many scientific studies, especially in epidemiology and environmental health, collecting individual measurements can be prohibitively expensive or require irreplaceable specimens. To overcome this, researchers often pool multiple samples together and measure the aggregate. While this saves resources, it means that instead of having individual response values, we only have the average response for each pool. This loss of information makes it challenging to use standard regression techniques that rely on individual data. This paper addresses the problem of estimating the relationship between a continuous response variable and covariates when only pooled response data are available, alongside individual-level covariate data.

Methods

The authors propose three novel local polynomial regression estimators for the conditional mean function, $m(x) = E(Y|X=x)$, specifically designed for pooled response data. These methods adapt the well-established local polynomial regression technique to handle pooled data under two common pooling designs: random pooling and homogeneous pooling.

  • Random Pooling: For this design, three estimators are developed:
  • Average-weighted estimator ($m_1(x)$): This estimator uses an average of individual-level weights to construct its objective function.
  • Product-weighted estimator ($m_2(x)$): This estimator uses a product of individual-level weights.
  • Marginal-integration estimator ($m_3(x)$): This estimator is inspired by the mean of a pseudo-response derived from the pooled data.
  • Homogeneous Pooling: For this design, the average-weighted and product-weighted estimators are adapted.

The paper then rigorously investigates the asymptotic properties (bias and variance) of these estimators under both pooling designs.

Datasets

The paper uses two real-life applications for illustration:

  1. Perfluorinated Chemicals (PFCs) Data: This dataset, from the National Health and Nutrition Examination Survey, relates the concentration levels of perfluorohexane sulfonic acid (PFHxS) in blood to an individual's age.
  2. Chemokines Data: This dataset, from the Collaborative Perinatal Project (CPP), examines the relationship between the concentration of GRO-a and MCP-3.

The simulation studies use synthetic data generated from four different data-generating processes (D1-D4) previously reported in Delaigle et al. (2009).

Key Insights and Ideas

  • The challenge of pooling: Pooling data, while cost-effective, introduces complexities in statistical modeling due to the loss of individual-level information.
  • Different pooling designs matter: The performance of the proposed estimators varies significantly depending on whether the pools are formed randomly or homogeneously.
  • Marginal-integration estimator shines under random pooling: The $m_3(x)$ estimator demonstrates superior performance under random pooling, showing good accuracy and consistency, especially compared to the other two proposed estimators.
  • Homogeneous pooling offers advantages: When data are homogeneously pooled, estimators $m_1(x)$ and $m_2(x)$ perform comparably to the individual-level estimator, suggesting that homogeneous pooling can preserve more information.
  • Curse of dimensionality in product-weighting: The product-weighted estimator ($m_2(x)$) suffers from high variability under random pooling, particularly as pool sizes increase, illustrating a form of the curse of dimensionality.
  • Bandwidth selection is crucial: The paper develops leave-one-out cross-validation procedures tailored for pooled data to select appropriate bandwidths for the estimators.

Key Results

  • Theoretical findings: The paper provides theoretical guarantees on the asymptotic properties of the proposed estimators, including their bias and variance under different pooling designs. For instance, Theorem 4.1 shows that the marginal-integration estimator ($m_3(x)$) is consistent under random pooling, while the average-weighted estimator ($m_1(x)$) is inconsistent. Theorem 4.2 indicates that both $m_1(x)$ and $m_2(x)$ are consistent under homogeneous pooling.
  • Simulation results: Simulation studies confirm the theoretical findings. Under random pooling, $m_3(x)$ generally outperforms $m_1(x)$ and $m_2(x)$ in terms of accuracy, while $m_2(x)$ exhibits higher variability. Under homogeneous pooling, $m_1(x)$ and $m_2(x)$ perform well, similar to the individual-level estimator.
  • Real-life applications: The application to PFCs data suggests that $m_3(x)$ captures the underlying relationship between PFHxS levels and age more accurately under random pooling, while $m_1(x)$ and $m_2(x)$ are comparable to the individual-level estimate under homogeneous pooling. The chemokine data analysis shows similar trends.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the proposed estimators might not fully exploit individual covariate information within a pool, suggesting room for improvement by more selective incorporation.
  • The analysis assumes that covariates are measured precisely. Future work could address situations where covariates are prone to measurement error.
  • Future Work:
    • Develop methods that more selectively incorporate individual covariate information within a pool.
  • Extend the methodology to handle pooled response data with individual-level covariates that are subject to measurement error.

Summary updated

Abstract

We propose local polynomial estimators for the conditional mean of a continuous response when only pooled response data are collected under different pooling designs. Asymptotic properties of these estimators are investigated and compared. Extensive simulation studies are carried out to compare finite sample performance of the proposed estimators under various model settings and pooling strategies. We apply the proposed local polynomial regression methods to two real-life applications to illustrate practical implementation and performance of the estimators for the mean function.
BibTeX
@article{wang2020local,
  author = {Dewei Wang and Xichen Mou and Xiang Li and Xianzheng Huang},
  title = {{Local polynomial regression for pooled response data}},
  journal = {Journal of Nonparametric Statistics},
  year = {2020},
  doi = {10.1080/10485252.2020.1834104},
}

Loading related content...

Figure