Volume 12, Issue s1 - Special Issue: Causal Inference; Issue Editors: Hélène Karcher and Fan Li

Objectives Specifying causal models to assess relationships among metal mixtures and cardiometabolic outcomes requires evidence-based models of the causal structures; however, such models have not been previously published. The objective of this study was to develop and evaluate a directed acyclic graph (DAG) diagraming metal mixture exposure and cardiometabolic outcomes. Methods We conducted a literature search to develop the DAG of metal mixtures and cardiometabolic outcomes. To evaluate consistency of the DAG, we tested the suggested conditional independence statements using linear and logistic regression analyses with data from the San Luis Valley Diabetes Study (SLVDS; n=1795). We calculated the proportion of statements supported by the data and compared this to the proportion of conditional independence statements supported by 1,000 DAGs with the same structure but randomly permuted nodes. Next, we used our DAG to identify minimally sufficient adjustment sets needed to estimate the association between metal mixtures and cardiometabolic outcomes (i.e., cardiovascular disease, fasting glucose, and systolic blood pressure). We applied them to the SLVDS using Bayesian kernel machine regression, linear mixed effects, and Cox proportional hazards models. Results From the 42 articles included in the review, we developed an evidence-based DAG with 74 testable conditional independence statements (43 % supported by SLVDS data). We observed evidence for an association between As and Mn and fasting glucose. Conclusions We developed, tested, and applied an evidence-based approach to analyze associations between metal mixtures and cardiometabolic health.

Objectives Propensity score (PS) weighting methods are commonly used to adjust for confounding in observational treatment comparisons. However, in the setting of substantial covariate imbalance, PS values may approach 0 and 1, yielding extreme weights and inflated variance of the estimated treatment effect. Adaptations of the standard inverse probability of treatment weights (IPTW) can reduce the influence of extremes, including trimming methods that exclude people with PS values near 0 or 1. Alternatively, overlap weighting (OW) optimizes criteria related to bias and variance, and performs well compared to other PS weighting and matching methods. However, it has not been compared to propensity score stratification (PSS). PSS has some of the same potential advantages; being insensitive extreme values. We sought to compare these methods in the setting of substantial covariate imbalance to generate practical recommendations. Methods Analytical derivations were used to establish connections between methods, and simulation studies were conducted to assess bias and variance of alternative methods. Results We find that OW is generally superior, particularly as covariate imbalance increases. In addition, a common method for implementing PSS based on Mantel–Haenszel weights (PSS-MH) is equivalent to a coarsened version of OW and can perform nearly as well. Finally, trimming methods increase bias across methods (IPTW, PSS and PSS-MH) unless the PS model is re-fit to the trimmed sample and weights or strata are re-derived. After trimming with re-fitting, all methods perform similarly to OW. Conclusions These results may guide the selection, implementation and reporting of PS methods for observational studies with substantial covariate imbalance.

In epidemiological studies, the capture-recapture (CRC) method is a powerful tool that can be used to estimate the number of diseased cases or potentially disease prevalence based on data from overlapping surveillance systems. Estimators derived from log-linear models are widely applied by epidemiologists when analyzing CRC data. The popularity of the log-linear model framework is largely associated with its accessibility and the fact that interaction terms can allow for certain types of dependency among data streams. In this work, we shed new light on significant pitfalls associated with the log-linear model framework in the context of CRC using real data examples and simulation studies. First, we demonstrate that the log-linear model paradigm is highly exclusionary. That is, it can exclude, by design, many possible estimates that are potentially consistent with the observed data. Second, we clarify the ways in which regularly used model selection metrics (e.g., information criteria) are fundamentally deceiving in the effort to select a “best” model in this setting. By focusing attention on these important cautionary points and on the fundamental untestable dependency assumption made when fitting a log-linear model to CRC data, we hope to improve the quality of and transparency associated with subsequent surveillance-based CRC estimates of case counts.