TL;DR
This paper introduces a novel approach, Generalized Linear Mixed Models with Natural Cubic Splines (GLMM+NS), for analyzing time-series data in air pollution health studies. By treating spline smoothing parameters as random effects, GLMM+NS offers a more flexible and potentially less biased way to model complex temporal trends compared to traditional Generalized Additive Models (GAMs), especially when pollution effects are small.
What Problem Does This Paper Address?
In environmental epidemiology, understanding the health impacts of air pollution often involves analyzing time-series data, like daily hospital admissions or mortality rates, in relation to daily pollutant levels. A common challenge is accounting for various temporal patterns that can influence health outcomes, such as long-term trends, seasonality, and day-of-week effects. Generalized Additive Models (GAMs) with natural cubic splines (NS) are widely used for this, allowing flexible "smoothing" of these time-varying factors. However, a key issue is determining the optimal degree of smoothing. Traditional methods, like using AIC or BIC, aim to optimize overall prediction but can lead to biased estimates of the air pollution effect itself. Furthermore, these methods often assume a fixed degree of smoothing throughout the entire study period, which might not accurately reflect the dynamic nature of real-world data. This paper addresses this by proposing a method that allows the smoothing to adapt more naturally over time.
Methods
The core innovation is the application of Generalized Linear Mixed Models (GLMMs) to spline smoothing in time-series analysis, termed GLMM+NS. Instead of treating the smoothing parameters of natural cubic splines as fixed (as in GLM+NS), they are modeled as random effects with an appropriate distribution. This allows the degree of smoothing to vary across the time series, potentially capturing local variations more effectively. The paper outlines the mathematical formulation of GLMM+NS, building upon the basis function representation of splines. For parameter estimation, they employ a penalized quasi-likelihood (PQL) approach, which is a common method for fitting GLMMs, implemented using the glmmPQL function in R. They also conduct simulation studies to compare GLMM+NS with the standard GLM+NS approach under various scenarios of pollution effect magnitudes and variance components.
Datasets
The primary dataset used for illustration is the Allegheny County Air Pollution Study (ACAPS). This dataset contains daily counts of cardiopulmonary hospital admissions for individuals over 65 years old, along with daily meteorological data (mean temperature) and daily ambient air levels of PM10 from 1995 to 2000. The data was collected from Allegheny County, Pennsylvania. The authors note that only a small number of data points were missing and were handled by exclusion.
Key Insights and Ideas
- Flexible Smoothing is Crucial: The paper highlights that assuming a fixed degree of smoothing for time-varying covariates might be unrealistic and could lead to biased estimates of air pollution effects. GLMM+NS offers a more adaptive approach.
- Random Effects for Smoothing Parameters: The core idea of treating spline smoothing parameters as random effects is a powerful way to introduce flexibility and account for uncertainty in the smoothing process.
- Trade-off in Small Effects: For very small true pollution effects, GLMM+NS shows similar absolute biases to GLM+NS but provides larger empirical standard deviations. This is interpreted as a more realistic reflection of uncertainty when the signal is weak.
- GLMM+NS as a Conservative Approach: The authors suggest that GLMM+NS provides a more conservative way to handle air pollution studies with time-varying covariates, especially when their true variations are uncertain.
Key Results
- Simulation Studies:
- For small true pollution effects (e.g., ≤ 0.01), GLMM+NS and GLM+NS yielded similar absolute biases.
- GLMM+NS consistently produced larger empirical standard deviations for the estimated pollution effect compared to GLM+NS. This is attributed to the inherent variability introduced by modeling smoothing parameters as random.
- The magnitude of the true pollution effect had a greater impact on parameter estimates than the variance of the random spline coefficients, particularly when the true effect was small.
- Illustrative Example (ACAPS data):
- The GLMM+NS model yielded a larger estimate for the effect of same-day PM10 on hospital admissions compared to the GLM+NS model.
- The standard errors for the PM10 effect were also larger in the GLMM+NS model.
- Root Mean Squared Errors (RMSE) for goodness-of-fit were slightly higher for GLMM+NS compared to GLM+NS, suggesting that the simpler GLM+NS might offer a slightly better overall fit in this specific case, though the differences were minimal.
Limitations and Future Work
- PQL Estimation: The paper acknowledges that the penalized quasi-likelihood (PQL) method used for fitting GLMM+NS, while computationally efficient, is an approximation and can lead to biased estimates of variance components. They compared it to SAS PROC NLMIXED, which uses Gaussian quadrature, and found comparable results, suggesting PQL is a reasonable choice for this application. However, they suggest exploring alternative estimation approaches for GLMMs.
- Complexity: Incorporating multiple smooth functions in GLMM+NS increases model complexity, which can sometimes lead to less stable estimates, as observed in one of their simulation scenarios (Table 3).
- Future Work: The authors implicitly suggest that further exploration of different estimation methods for GLMMs would be beneficial. They also highlight the general challenge of model selection and the trade-offs between model complexity and estimation precision in time-series air pollution studies.
Summary updated