Data-driven Covariate Selection for Confounding Adjustment by Focusing on the Stability of the Effect Estimator

Valid inference of cause-and-effect relations in observational studies necessitates adjusting for common causes of the focal predictor (i.e., treatment) and the outcome. When such common causes, henceforth termed confounders, remain unadjusted for, they generate spurious correlations that lead to biased causal effect estimates. But routine adjustment for all available covariates, when only a subset are truly confounders, is known to yield potentially inefficient and unstable estimators. In this article, we introduce a data-driven confounder selection strategy that focuses on stable estimation of the treatment effect. The approach exploits the causal knowledge that after adjusting for confounders to eliminate all confounding biases, adding any remaining non-confounding covariates associated with only treatment or outcome, but not both, should not systematically change the effect estimator. The strategy proceeds in two steps. First, we prioritize covariates for adjustment by probing how strongly each covariate is associated with treatment and outcome. Next, we gauge the stability of the effect estimator by evaluating its trajectory adjusting for different covariate subsets. The smallest subset that yields a stable effect estimate is then selected. Thus, the strategy offers direct insight into the (in)sensitivity of the effect estimator to the chosen covariates for adjustment. The ability to correctly select confounders and yield valid causal inference following data-driven covariate selection is evaluated empirically using extensive simulation studies. Furthermore, we compare the proposed method empirically with routine variable selection methods. Finally, we demonstrate the procedure using two publicly available real-world datasets.

Data-driven Covariate Selection for Confounding Adjustment by Focusing on the Stability of the Effect Estimator

Valid inference of cause-and-effect relations in observational studies necessitates adjusting for common causes of the focal predictor (i.e., treatment) and the outcome. When such common causes, henceforth termed confounders, remain unadjusted for, they generate spurious correlations that lead to biased causal effect estimates. But routine adjustment for all available covariates, when only a subset are truly confounders, is known to yield potentially inefficient and unstable estimators. In this article, we introduce a data-driven confounder selection strategy that focuses on stable estimation of the treatment effect. The approach exploits the causal knowledge that after adjusting for confounders to eliminate all confounding biases, adding any remaining non-confounding covariates associated with only treatment or outcome, but not both, should not systematically change the effect estimator. The strategy proceeds in two steps. First, we prioritize covariates for adjustment by probing how strongly each covariate is associated with treatment and outcome. Next, we gauge the stability of the effect estimator by evaluating its trajectory adjusting for different covariate subsets. The smallest subset that yields a stable effect estimate is then selected. Thus, the strategy offers direct insight into the (in)sensitivity of the effect estimator to the chosen covariates for adjustment. The ability to correctly select confounders and yield valid causal inference following data-driven covariate selection is evaluated empirically using extensive simulation studies. Furthermore, we compare the proposed method empirically with routine variable selection methods. Finally, we demonstrate the procedure using two publicly available real-world datasets.

Differential Impacts of COVID-19 Lockdowns on PM2.5 across the United States

ABSTRACTThe COVID-19 pandemic has induced large-scale social, economic, and behavioral changes, presenting a unique opportunity to study how air pollution is affected by unprecedented societal shifts. At each of 455 PM2.5 monitoring sites across the United States, we conduct a causal inference analysis to determine the impacts of COVID-19 interventions and behavioral changes (“lockdowns”) on PM2.5 concentrations. Our approach allows for rigorous confounding adjustment and provides highly spatio-temporally resolved effect estimates. We find that, with the exception of the Southwest, most of the US experienced increases in PM2.5 during lockdown, compared to the concentrations expected under business-as-usual. To investigate possible drivers of this phenomenon, we use regression to characterize the relationship of many environmental, geographical, meteorological, mobility, and socioeconomic factors with the lockdown-attributable changes in PM2.5. Our findings have immense environmental policy relevance, suggesting that large-scale mobility and economic activity reductions may be insufficient to substantially and uniformly reduce PM2.5.

Further investigation of gateway effects using the PATH study

Background: Interest exists in whether youth e-cigarette use (“vaping”) increases risk of initiating cigarette smoking. Using Waves 1 and 2 of the US PATH study we previously reported adjustment for vaping propensity using Wave 1 variables explained about 80% of the unadjusted relationship. Here data from Waves 1 to 3 are used to avoid over-adjustment if Wave 1 vaping affected variables recorded then. Methods: Main analyses M1 and M2 concerned Wave 2 never smokers who never vaped by Wave 1, linking Wave 2 vaping to Wave 3 smoking initiation, adjusting for predictors of vaping based on Wave 1 data using differing propensity indices. M3 was similar but derived the index from Wave 2 data. Sensitivity analyses excluded Wave 1 other tobacco product users, included other product use as another predictor, or considered propensity for smoking or any tobacco use, not vaping. Alternative analyses used exact age (not previously available) as a confounder not grouped age, attempted residual confounding adjustment by modifying predictor values using data recorded later, or considered interactions with age. Results: In M1, adjustment removed about half the excess OR (i.e. OR–1), the unadjusted OR, 5.60 (95% CI 4.52-6.93), becoming 3.37 (2.65-4.28), 3.11 (2.47-3.92) or 3.27 (2.57-4.16), depending whether adjustment was for propensity as a continuous variable, as quintiles, or the variables making up the propensity score. Many factors had little effect: using grouped or exact age; considering other products; including interactions; or using predictors of smoking or tobacco use rather than vaping. The clearest conclusion was that analyses avoiding over-adjustment explained about half the excess OR, whereas analyses subject to over-adjustment explained about 80%. Conclusions: Although much of the unadjusted gateway effect results from confounding, we provide stronger evidence than previously of some causal effect of vaping, though doubts still remain about the completeness of adjustment.

A Systematic Review of the Association Between Psychological Stress and Dementia Risk in Humans

Background: It has been estimated that one third of dementia cases may be preventable through modifiable lifestyle interventions. Epidemiological evidence suggests a link between stressful life events and aging-related cognitive decline and dementia; however, inherent methodological limitations in examining subjective and biological measures of stress separately leads to interpretive constraints. Objective: The aim of the current study was to conduct a systematic review of the research literature investigating the effect of perceived and biological measures of stress on dementia risk. Methods: A systematic review was conducted of cohort, case-control, longitudinal prospective or retrospective studies examining the association between stress and risk of developing dementia. Studies were identified from a systematic search across major electronic databases from inception to February 2020. Results: Overall, 22 studies were identified including a total of 496,556 participants, approximately 50% were females, with sample sizes ranging from 62–270,977. There was considerable heterogeneity in the definition and measurement of stress. Most of the identified studies reported a significant positive association between stress and dementia risk. Conclusion: Evidenced from the current review is that personality traits linked to increased perceived stress and elevated reported perceived stress, are associated with greater statistical risk for dementia. However, this review highlights that caution must be exhibited in interpreting these findings, as methodological issues with confounding adjustment may mediate these results. Future research should focus on the investigation of stress on dementia risk with a full range of confounding adjustment, and on biological measures of stress.

Long-term exposure to air-pollution and COVID-19 mortality in England: a hierarchical spatial analysis

Background: Recent studies suggested a link between long-term exposure to air-pollution and COVID-19 mortality. However, due to their ecological design, based on large spatial units, they neglect the strong localised air-pollution patterns, and potentially lead to inadequate confounding adjustment. We investigated the effect of long-term exposure to NO2 and PM2.5 on COVID-19 deaths up to June 30, 2020 in England using high geographical resolution. Methods: We included 38 573 COVID-19 deaths up to June 30, 2020 at the Lower Layer Super Output Area level in England (n=32 844 small areas). We retrieved averaged NO2 and PM2.5 concentration during 2014-2018 from the Pollution Climate Mapping. We used Bayesian hierarchical models to quantify the effect of air-pollution while adjusting for a series of confounding and spatial autocorrelation. Findings: We find a 0.5% (95% credible interval: -0.2%-1.2%) and 1.4% (-2.1%-5.1%) increase in COVID-19 mortality rate for every 1μg/m3 increase in NO2 and PM2.5 respectively, after adjusting for confounding and spatial autocorrelation. This corresponds to a posterior probability of a positive effect of 0.93 and 0.78 respectively. The spatial relative risk at LSOA level revealed strong patterns, similar for the different pollutants. This potentially captures the spread of the disease during the first wave of the epidemic. Interpretation: Our study provides some evidence of an effect of long-term NO2 exposure on COVID-19 mortality, while the effect of PM2.5 remains more uncertain. Funding: Medical Research Council, Wellcome Trust, Environmental Protection Agency and National Institutes of Health.

A Systematic Review of Methods Used for Confounding Adjustment in Observational Economic Evaluations in Cardiology Conducted between 2013 and 2017

Background. Observational economic evaluations (i.e., economic evaluations in which treatment allocation is not randomized) are prone to confounding bias. Prior reviews published in 2013 have shown that adjusting for confounding is poorly done, if done at all. Although these reviews raised awareness on the issues, it is unclear if their results improved the methodological quality of future work. We therefore aimed to investigate whether and how confounding was accounted for in recently published observational economic evaluations in the field of cardiology. Methods. We performed a systematic review of PubMed, Embase, Cochrane Library, Web of Science, and PsycInfo databases using a set of Medical Subject Headings and keywords covering topics in “observational economic evaluations in health within humans” and “cardiovascular diseases.” Any study published in either English or French between January 1, 2013, and December 31, 2017, addressing our search criteria was eligible for inclusion in our review. Our protocol was registered with PROSPERO (CRD42018112391). Results. Forty-two (0.6%) out of 7523 unique citations met our inclusion criteria. Fewer than half of the selected studies adjusted for confounding ( n = 19 [45.2%]). Of those that adjusted for confounding, propensity score matching ( n = 8 [42.1%]) and other matching-based approaches were favored ( n = 8 [42.1%]). Our results also highlighted that most authors who adjusted for confounding rarely justified their methodological choices. Conclusion. Our results indicate that adjustment for confounding is often ignored when conducting an observational economic evaluation. Continued knowledge translation efforts aimed at improving researchers’ knowledge regarding confounding bias and methods aimed at addressing this issue are required and should be supported by journal editors.

Further investigation of gateway effects using the PATH study

Background: Interest exists in whether youth e-cigarette use (“vaping”) increases risk of initiating cigarette smoking. Using Waves 1 and 2 of the US PATH study we reported that adjustment for vaping propensity using Wave 1 variables explained about 80% of the unadjusted relationship. Here we use data from Waves 1 to 3 to avoid over-adjustment if Wave 1 vaping affected variables recorded then. Methods: Our main analysis M1 concerned Wave 2 never smokers who never vaped by Wave 1, linking Wave 2 vaping to Wave 3 smoking initiation, adjusting for Wave 1 predictors. We conducted sensitivity analyses that: excluded Wave 1 other tobacco product users; included other product use as an extra predictor; or considered propensity for smoking or any tobacco use, rather than vaping. We also conducted analyses that: adjusted for propensity as derived originally; ignored Wave 1 data; used exact age (not previously available) as a confounder rather than grouped age; attempted residual confounding adjustment by modifying predictor values using data recorded later; or considered interactions with age. Results: In M1, adjustment removed about half the excess OR (i.e. OR–1), the unadjusted OR, 5.60 (95% CI 4.52-6.93), becoming 3.37 (2.65-4.28), 3.11 (2.47-3.92) or 3.27 (2.57-4.16), depending whether adjustment was for propensity as a continuous variable, as quintiles, or for the variables making up the propensity score. Many factors had little effect: using grouped or exact age; considering other products; including interactions; or using predictors of smoking or tobacco use rather than vaping. The clearest conclusion was that analyses avoiding over-adjustment explained about half the excess OR, whereas analyses subject to over-adjustment explained about 80%. Conclusions: Although much of the unadjusted gateway effect results from confounding, we provide stronger evidence than previously of some causal effect of vaping, though some doubts still remain about the completeness of adjustment.