Methods/Statistics
Partial Effects in Environmental Mixtures – Guidance on Methods and Implications Alex Keil* Maria Kamenetsky Barrett M. Welch Paige A. Bommarito Jessie P. Buckley Katie M. O’Brien Alexandra J. White Thomas F. McElrath David E. Cantonwine Kelly K. Ferguson Alexander P. Keil
Exposure mixtures such as water contamination are pervasive in the environment. Mixtures methods, such as quantile g-computation (QGC) and weighted quantile sums (WQS) regression, focus on the joint effects of increasing all components of the mixture simultaneously. There is also an interest in refining those estimands to target negative and positive “partial effects”, which estimate impacts of increasing only a subset of the mixture. While the performance of QGC and WQS regression have been evaluated for how well they estimate joint effects, their performance for estimating partial effects is unknown. We study the performance of QGC and WQS and their data-adaptive extensions (sample-splitting, model-averaging, penalized regression, no sample-splitting) in estimating partial effects. We contrast these methods with an approach based on a priori knowledge (QGCAP), that determines the partition of negative and positive exposures using subject-matter expertise. In simulations, we compare performance across methods and assess the impact of four study characteristics on performance: 1) exposure correlation, 2) sample size, 3) spread of partial effect across more exposures, and 4) imbalance in negative and positive effects. Only accurate a priori knowledge of exposures guaranteed precise estimation by QGC, measured by mean squared error (MSE). Other methods demonstrated some bias in nearly every setting (Figure 1). Error and bias increased as exposure correlation increased, sample sizes shrunk, as the partial effect was spread across more exposures, and as imbalance between negative and positive effects increased. QGC-based methods yielded more predictable bias and were the only approximately unbiased methods at large sample sizes. Outside of QGCAP, no method for estimating partial effects was best in all scenarios. We provide guidance for practitioners on the use and limitations of statistical approaches for disentangling harmful and helpful components of a mixture.
