Collider Stratification Bias Research Articles

Rethinking animal attrition in preclinical research: Expressing causal mechanisms of selection bias using directed acyclic graphs.

Animal attrition in preclinical experiments can introduce bias in the estimation of causal treatment effects, as the treatment-outcome association in surviving animals may not represent the causal effect of interest. This can compromise the internal validity of the study despite randomization at the outset. Directed Acyclic Graphs (DAGs) are useful tools to transparently visualize assumptions about the causal structure underlying observed data. By illustrating relationships between relevant variables, DAGs enable the detection of even less intuitive biases, and can thereby inform strategies for their mitigation. In this study, we present an illustrative causal model for preclinical stroke research, in which animal attrition induces a specific type of selection bias (i.e., collider stratification bias) due to the interplay of animal welfare, initial disease severity and negative side effects of treatment. Even when the treatment had no causal effect, our simulations revealed substantial bias across different scenarios. We show how researchers can detect and potentially mitigate this bias in the analysis phase, even when only data from surviving animals are available, if knowledge of the underlying causal process that gave rise to the data is available. Collider stratification bias should be a concern in preclinical animal studies with severe side effects and high post-randomization attrition.

Potential Biases in Test-Negative Design Studies of COVID-19 Vaccine Effectiveness Arising from the Inclusion of Asymptomatic Individuals.

The test-negative design (TND) is a popular method for evaluating vaccine effectiveness (VE). A "classical" TND study includes symptomatic individuals tested for the disease targeted by the vaccine to estimate VE against symptomatic infection. However, recent applications of the TND have attempted to estimate VE against infection by including all tested individuals, regardless of their symptoms. In this article, we use directed acyclic graphs and simulations to investigate potential biases in TND studies of COVID-19 VE arising from the use of this "alternative" approach, particularly when applied during periods of widespread testing. We show that the inclusion of asymptomatic individuals can potentially lead to collider stratification bias, uncontrolled confounding by health and healthcare-seeking behaviors (HSBs), and differential outcome misclassification. While our focus is on the COVID-19 setting, the issues discussed here may also be relevant in the context of other infectious diseases. This may be particularly true in scenarios where there is either a high baseline prevalence of infection, a strong correlation between HSBs and vaccination, different testing practices for vaccinated and unvaccinated individuals, or settings where both the vaccine under study attenuates symptoms of infection and diagnostic accuracy is modified by the presence of symptoms.

Multiple Sclerosis Risk Among Anti-tumor Necrosis Factor Alpha Users: A Methodological Review of Observational Studies Based on Real-world Data.

Epidemiologic studies on the risk of multiple sclerosis (MS) or demyelinating events associated with anti-tumor necrosis factor alpha (TNFα) use among patients with rheumatic diseases or inflammatory bowel diseases have shown conflicting results. Causal directed acyclic graphs (cDAGs) are useful tools for understanding the differing results and identifying the structure of potential contributing biases. Most of the available literature on cDAGs uses language that might be unfamiliar to clinicians. This article demonstrates how cDAGs can be used to determine whether there is a confounder, a mediator or collider-stratification bias and when to adjust for them appropriately. We also use a case study to show how to control for potential biases by drawing a cDAG depicting anti-TNFα use and its potential to contribute to MS onset. Finally, we describe potential biases that might have led to contradictory results in previous studies that examined the effect of anti-TNFα and MS, including confounding, confounding by contraindication, and bias due to measurement error. Clinicians and researchers should be cognizant of confounding, confounding by contraindication, and bias due to measurement error when reviewing future studies on the risk of MS or demyelinating events associated with anti-TNFα use. cDAGs are a useful tool for selecting variables and identifying the structure of different biases that can affect the validity of observational studies.

Investigating the “sex paradox” in pulmonary arterial hypertension: Results from the Pulmonary Hypertension Association Registry (PHAR)

BackgroundFemale sex is a significant risk factor for pulmonary arterial hypertension (PAH), yet males with PAH have worse survival – a phenomenon referred to as the “sex paradox” in PAH. MethodsAll adult PAH patients in the Pulmonary Hypertension Association Registry (PHAR) with congruent sex and gender were included. Baseline differences in demographics, hemodynamics, functional parameters, and quality of life were assessed by sex. Kaplan-Meier survival analysis was used to evaluate survival by sex. Mediation analysis was conducted with Cox proportional hazards regression by comparing the unadjusted hazard ratios for sex before and after adjustment for covariates. The plausibility of collider-stratification bias was assessed by modeling how large an unmeasured factor would have to be to generate the observed sex-based mortality differences. Subgroup analysis was performed on idiopathic and incident patients. ResultsAmong the 1,891 patients included, 75% were female. Compared to men, women had less favorable hemodynamics, lower six-minute walk distance, more PAH therapies, and worse functional class; however, sex-based differences were less pronounced when accounting for body surface area or expected variability by gender. On multivariate analysis, women had a 48% lower risk of death compared to men (Hazard Ratio 0.52, 95% Confidence interval 0.36 – 0.74, p<0.001). Modeling found that under reasonable assumptions collider-stratification could account for sex-based differences in mortality. ConclusionsIn this large registry of PAH patients new to a care center, men had worse survival than women despite having more favorable baseline characteristics. Collider-stratification bias could account for the observed greater mortality among men.

Selection Bias Requires Selection: The Case of Collider Stratification Bias.

In epidemiology, collider stratification bias, the bias resulting from conditioning on a common effect of two causes, is oftentimes considered a type of selection bias, regardless of the conditioning methods employed. In this commentary, we distinguish between two types of collider stratification bias: collider restriction bias due to restricting to one level of a collider (or a descendant of a collider) and collider adjustment bias through inclusion of a collider (or a descendant of a collider) in a regression model. We argue that categorizing collider adjustment bias as a form of selection bias may lead to semantic confusion, as adjustment for a collider in a regression model does not involve selecting a sample for analysis. Instead, we propose that collider adjustment bias can be better viewed as a type of overadjustment bias. We further provide two distinct causal diagram structures to distinguish collider restriction bias and collider adjustment bias. We hope that such a terminological distinction can facilitate easier and clearer communication.

Collider Stratification Bias I: Principles and Structure.

Collider Stratification Bias I: Principles and Structure.

Selection Bias in Health Research: Quantifying, Eliminating, or Exacerbating Health Disparities?

To summarize recent literature on selection bias in disparities research addressing either descriptive or causal questions, with examples from dementia research. Defining a clear estimand, including the target population, is essential to assess whether generalizability bias or collider-stratification bias are threats to inferences. Selection bias in disparities research can result from sampling strategies, differential inclusion pipelines, loss to follow-up, and competing events. If competing events occur, several potentially relevant estimands can be estimated under different assumptions, with different interpretations. The apparent magnitude of a disparity can differ substantially based on the chosen estimand. Both randomized and observational studies may misrepresent health disparities or heterogeneity in treatment effects if they are not based on a known sampling scheme. Researchers have recently made substantial progress in conceptualization and methods related to selection bias. This progress will improve the relevance of both descriptive and causal health disparities research.

Adjustment for collider bias in the hospitalized Covid-19 setting

BackgroundCausal directed acyclic graphs (cDAGs) are frequently used to identify confounding and collider bias. We demonstrate how to use causal directed acyclic graphs to adjust for collider bias in the hospitalized Covid-19 setting. Materials and methodsAccording to the cDAGs, three types of modeling have been performed. In model 1, only vaccination is entered as an independent variable. In model 2, in addition to vaccination, age is entered the model to adjust for collider bias due to the conditioning of hospitalization. In model 3, comorbidities are also included for adjustment of collider bias due to the conditioning of hospitalization in different biasing paths intercepting age and comorbidities. ResultsThere was no evidence of the effect of vaccination on preventing death due to Covid-19 in model 1. In the second model, where age was included as a covariate, a protective role for vaccination became evident. In model 3, after including chronic diseases as other covariates, the protective effect was slightly strengthened. ConclusionStudying hospitalized patients is subject to collider-stratification bias. Like confounding, this type of selection bias can be adjusted for by inclusion of the risk factors of the outcome which also affect hospitalization in the regression model.

Statins and Mortality in COPD: A Methodological Review of Observational Studies

Randomized controlled trials and observational studies have reported conflicting results on the potential beneficial effects of statins on mortality in patients with chronic obstructive pulmonary disease (COPD). We performed a systematic search of the literature to review all observational studies reporting relative risks of death with statin use in COPD, focusing on potential sources of bias. We identified 15 observational studies, out of 2835, of which 12 were affected by time-related and other biases and the remaining 3 by confounding bias. All 15 studies were also subject to confounding bias due to lack of adjustment for important COPD-related factors. The risk of death associated with statin use was reduced across all 15 studies (pooled relative risk (PRR) 0.66; 95% CI: 0.59-0.74). The reduction was observed in 7 studies with immortal time bias (PRR 0.62; 95%: 0.53-0.72), two with collider-stratification bias (PRR 0.60; 95% CI: 0.45-0.80), one with time-window bias (RR 0.61; 95% CI: 0.38-0.98), one with immeasurable time bias (RR 0.50; 95% CI: 0.40-0.62), and one with exposure misclassification (RR 0.86; 95% CI: 0.72-1.03). The three studies that avoided these biases were, however, affected by confounding bias resulting in a PRR of 0.77 (95% CI: 0.61-0.98). In conclusion, the observational studies investigating statin use and mortality in COPD are affected by major biases, many of which can result in spurious protective effects. Well-designed observational studies that carefully emulate randomized trials are needed to resolve this uncertainty regarding the potential beneficial benefits of statins on mortality in patients with COPD.

The impact of sociodemographic status on the association of classical cardiovascular risk factors with coronary artery disease: a stratified Mendelian randomization study

ObjectivesLow socioeconomic status (SES) is associated with cardiovascular risk factors and increased coronary artery disease (CAD) risk. We tested whether SES is an effect modifier of the association between classical cardiovascular risk factors and CAD using SES-stratified Mendelian Randomization in European-ancestry participants from UK Biobank. Study Design and SettingWe calculated weighted genetic risk scores (GRS) for the risk factors body mass index (BMI), systolic blood pressure, low-density lipoprotein cholesterol, and triglycerides. Participants were stratified by Townsend deprivation index score. Logistic regression models were used to investigate associations between GRSs and CAD occurrence. Additionally, stratification based on GRS-adjusted Townsend deprivation index residuals was conducted to correct for possible collider-stratification bias. ResultsIn a total sample size of N = 446,485, with 52,946 cases, the risk for CAD per standard deviation increase in genetically influenced BMI was highest in the group with the lowest 25% SES (odds ratio: 1.126, 95% confidence interval: 1.106–1.145; odds ratio: 1.081, 95% confidence interval: 1.059–1.103 in high SES), remaining similar after controlling for possible collider-stratification bias. The effects of genetically influenced systolic blood pressure, low-density lipoprotein cholesterol, and triglyceride on CAD were similar between SES groups. ConclusionCAD risk attributable to increased BMI is not homogenous and could be modified by SES. This emphasizes the need of tailor-made approaches for BMI-associated CAD risk reduction.

Adverse childhood experiences and adult outcomes using a causal framework perspective: Challenges and opportunities

BackgroundResearch on the effect of adverse childhood experiences (ACEs) on adult outcomes has typically relied on retrospective assessment of ACEs and cumulative scores. However, this approach raises methodological challenges that can limit the validity of findings. ObjectiveThe aims of this paper are 1) to present the value of directed acyclic graphs (DAGs) to identify and mitigate potential problems related to confounding and selection bias, and 2) to question the meaning of a cumulative ACE score. ResultsAdjusting for variables that post-date childhood could block mediated pathways that are part of the total causal effect while conditioning on adult variables, which often serve as proxies for childhood variables, can create collider stratification bias. Because exposure to ACEs can affect the likelihood of reaching adulthood or study entry, selection bias could be introduced via restricting selection on a variable affected by ACEs in the presence of unmeasured confounding. In addition to challenges regarding causal structure, using a cumulative score of ACEs assumes that each type of adversity will have the same effect on a given outcome, which is unlikely considering differing risk across adverse experiences. ConclusionsDAGs provide a transparent approach of the researchers' assumed causal relationships and can be used to overcome issues related to confounding and selection bias. Researchers should be explicit about their operationalization of ACEs and how it is to be interpreted in the context of the research question they are trying to answer.

Preferred Language Mediates Association Between Race, Ethnicity, and Delayed Presentation in Critically Ill Patients With COVID-19.

We hypothesized that preferred language mediates the association between race, ethnicity and delays to care. Multicenter, retrospective cohort study of adults with COVID-19 consecutively admitted to the ICU in three Massachusetts hospitals in 2020. Causal mediation analysis was performed to evaluate potential mediators including preferred language, insurance status, and neighborhood characteristics. Non-Hispanic White (NHW) patients (157/442, 36%) were more likely to speak English as their preferred language (78% vs. 13%), were less likely to be un- or under-insured (1% vs. 28%), lived in neighborhoods with lower social vulnerability index (SVI) than patients from racial and ethnic minority groups (SVI percentile 59 [28] vs. 74 [21]) but had more comorbidities (Charlson comorbidity index 4.6 [2.5] vs. 3.0 [2.5]), and were older (70 [13.2] vs. 58 [15.1] years). From symptom onset, NHW patients were admitted 1.67 [0.71-2.63] days earlier than patients from racial and ethnic minority groups (p < 0.01). Non-English preferred language was associated with delay to admission of 1.29 [0.40-2.18] days (p < 0.01). Preferred language mediated 63% of the total effect (p = 0.02) between race, ethnicity and days from symptom onset to hospital admission. Insurance status, social vulnerability, and distance to the hospital were not on the causal pathway between race, ethnicity and delay to admission. Preferred language mediates the association between race, ethnicity and delays to presentation for critically ill patients with COVID-19, although our results are limited by possible collider stratification bias. Effective COVID-19 treatments require early diagnosis, and delays are associated with increased mortality. Further research on the role preferred language plays in racial and ethnic disparities may identify effective solutions for equitable care.

#3559 THE IMPACT OF OBESITY ON CLINICAL OUTCOMES IN VARIOUS LEVELS OF CO-MORBIDITY OF NON-DIALYSIS CHRONIC KIDNEY DISEASE PATIENTS

Abstract Background and Aims Obesity is a major issue with an estimated prevalence of 1.9 billion adults worldwide. Obesity is an important risk factor for premature death and the development of non-communicable diseases such as diabetes mellitus (DM), heart diseases, and chronic kidney disease (CKD). However, mounting evidence in the literature describes a reverse association whereby obesity may have a protective effect on mortality; this is sometimes referred to as the “obesity paradox”. Several reports question the concept of obesity paradox claiming methodology flaws such as collider stratification bias. In this study, we aimed to examine the effects of obesity on the combined outcomes of all-cause mortality (ACM) and renal replacement therapy (RRT) incidence in a cohort of patients with non-dialysis dependent CKD (NDD-CKD) by correcting for major risk factors to reduce the risk of bias. Method This retrospective study was undertaken on all patients with a documented body mass index (BMI) in the Salford Kidney Study database from October 2002 until December 2016. Patients were grouped according to their BMI into normal weight [BMI 18.5-24.9 kg/m2], overweight [BMI 25–29.9 kg/m2 and obese [BMI&amp;gt; 30 kg/m2]. Patients were also grouped according to their level of co-morbidity into 4 groups: group 1 had CKD only; group 2 had CKD and heart failure (HF); group 3 had CKD and DM; and group 4 had CKD, DM, and HF. Univariate Cox regression as well as three stepwise models of multivariate analysis were performed to study the strength of association between BMI categories and combined outcomes (incidence of RRT and ACM) across the 4 groups of different clusters of co-morbidity. Results A total of 2416 patients were included in the analysis. The median age of the cohort was 67.3 years [IQR 55.9-75.6], 61.8% were male, and 96.4% were of white ethnicity. The median BMI was 28.1 kg/m2 [IQR 24.7-32.6] and the median estimated glomerular filtration rate (eGFR) was 30.7 ml/min/1.73m2 [IQR 20.4-43.5]. At baseline, patients with increasing level of co-morbidity tended to be older with higher prevalence of hypertension (HTN), angina, myocardial infarction (MI), and stroke with lower baseline eGFR. The risk of combined outcomes followed the same trend in the three BMI groups, risk is higher with higher index of co-morbidity (p &amp;lt;0.001). Further analysis of four subgroups of co-morbidity was undertaken. A univariate Cox regression analysis for group 1 [CKD only, n = 1351], and group 2 [CKD and HF, n = 227] showed that patients with obesity had significant lower rates of combined outcomes compared to patients with normal BMI (HR 0.75; 95%CI = 0.63-0.89; p = 0.001 and HR 0.56; 95%CI = 0.38-0.82; p = 0.003 for group 1 and group 2 respectively). In multivariate models, obesity consistently proved to be a strong protective factor against combined outcomes (HR 0.77; 95%CI = 0.65-0.92; p = 0.005 for group 1 and HR 0.53; 95%CI = 0.34-0.83; p = 0.005 for group 2). This was independent of age, gender, HTN, angina, stroke, MI, and prescription of statins and angiotensin converting enzyme inhibitors. For group 3 [CKD and DM, n = 614], and group 4 [CKD, DM, and HF, n = 190], there was no significant difference in the combined outcomes between the different BMI groups when using univariate Cox regression analysis (for patients with obesity: HR 0.78; 95%CI = 0.61-1.01; p = 0.060 and HR 0.70; 95%CI = 0.43-1.16; p = 0.166 for both groups respectively). There was no significant difference in the incidence of RRT in any of the four groups. Conclusion In our largely white NDD-CKD cohort of patients, there was evidence of increasing risk of RRT or ACM as comorbidity increased irrespective of BMI. This is not surprising as ACM would be expected to increase as the burden of disease increases. However, when comparing the effect of BMI within groups, obesity was protective against combined outcomes in group 1 (CKD only) and group 2 (CKD+HF). This ‘protective’ effect was not seen in patients who had concomitant diabetes. These data suggest that diabetes is a potent predictor of outcomes irrespective of BMI, however, in patients without diabetes, obesity may play a protective role.

Invited Commentary: Modern Epidemiology Confronts COVID-19-Reflections From Psychiatric Epidemiology.

Dimitris et al. (Am J Epidemiol. 2022;191(6):980-986) outline how the coronavirus disease 2019 (COVID-19) pandemic has, with mixed results, put epidemiology under the spotlight. While epidemiologic theory and methods have been critical in many successes, the ongoing global death toll from severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) and the sometimes chaotic public messaging underscore that epidemiology as a field has room for improvement. Here, we use examples from psychiatric epidemiologic studies conducted during the COVID-19 era to reflect on errors driven by overlooking specific major methodological advances of modern epidemiology. We focus on: 1) use of nonrepresentative sampling in online surveys, which limits the potential knowledge to be gained from descriptive studies and amplifies collider stratification bias in causal studies; and 2) failure to acknowledge multiple versions of exposures (e.g., lockdown, school closure) and differences in prevalence of effect measure modifiers across contexts, which causes violations of the consistency assumption and lack of effect transportability. We finish by highlighting: 1) the heterogeneity of psychiatric epidemiologic results during the pandemic across place and sociodemographic groups and over time; 2) the importance of following the foundational advancements of modern epidemiology even in emergency settings; and 3) the need to limit the role of political agendas in cherry-picking and reporting epidemiologic evidence.

The effect of BMI at cancer diagnosis on survival of patients with head and neck carcinoma

Abstract Background Previous findings suggest a positive association between body mass index (BMI) and survival from head and neck cancer (HNC). The aim of this study is to investigate the prognostic role of BMI at the time of diagnosis in a large international cohort of HNC patients. Methods We performed a pooled analysis of studies included in the INHANCE consortium. Cases were adults with HNCs of the oral cavity, oropharynx, hypopharynx, and larynx. We used Cox proportional hazards models to estimate the adjusted hazard ratios (HR) for overall survival and HNC-specific survival, by cancer site. Subgroups analyses were performed according to smoking status and duration of follow-up. Results The study included 10,177 patients from 10 studies worldwide with a median follow-up of 48 months; 3654 patients (35.9%) died from all causes, including 1202 (11.8%) from HNC. Underweight patients had lower overall survival (HR = 1.69, 95% CI: 1.31-2.19) respect to normal BMI patients (BMI=18.5-24.9 kg/m2) with consistent results across the HNC sites. In HNC-specific mortality analyses, the survival for underweight patients was not significant, except for underweight patients with oropharyngeal cancer (HR = 1.43, 95% CI: 1.11-1.83). Overweight and obese patients for oropharyngeal cancers had a favourable HNC-specific survival (HR = 0.50 (95% CI: 0.33-0.75) and HR = 0.51 (95% CI: 0.36-0.72), respectively). Among never smokers, overall BMI status was not associated with HNC-specific survival. Among ever smokers overweight and obese categories showed a favourable HNC-specific survival (HR = 0.69 (95% CI: 0.56-0.86) and HR = 0.70 (95% CI: 0.61-0.80)). Conclusions Our findings show that high BMI values increase the survival rates in smoking patients with HNC, suggesting that a nutritional reserve may help patients survive HNC cancer. This effect, however, may be partly explained by residual confounding, reverse causation, and collider stratification bias. Key messages • Our analysis reports the results of the largest available pooled analysis on the prognostic significance of BMI in the survival of 10,177 HNC patients from 10 studies worldwide. • Lower overall survival was observed in underweight patients for all HNC sites and a lower HNC-specific survival was observed with oropharyngeal cancer. Adiposity could serve as a nutrient reserve.

Relationship of obesity to mortality in 47,399 patients from nationwide LIPIDOGRAM 2004–2015 cohort studies

Abstract Background Evidence suggests that obesity may be associated with lower mortality rates in various cohorts of patients (obesity paradox). One possible explanation is unmeasured confounding factors induced by selection bias (collider stratification bias). Therefore, data from large cohorts with few exclusion criteria are of value as they may minimize this effect. Purpose To assess the association between two obesity measures body mass index (BMI; kg/m2) and waist circumference (WC; cm) with mortality in nationwide cohort of 47,399 patients. Methods Lipidogram studies were carried out in Poland in 2004, 2006 and 2015 in the population representative for patients in primary care setting. The median follow up was 5570 days. Patients were recruited in all 16 administrative regions in Poland and physicians were proportionally to the number of inhabitants in a given administrative region. Each patient was given a questionnaire on chronic diseases, treatment and lifestyle. Questionnaire was administered by physician. Association between BMI, WC and mortality was assessed by multivariate Cox regression and spline regression models. Weight status was classified as underweight (BMI &amp;lt;18.5), normal weight (BMI 18.5–24.9), overweight (BMI 25–29.9), class I obesity (BMI 30–34.9), class II obesity BMI (35–39.9), and class III obesity BMI (&amp;gt;40). Results The prevalence of hypertension, dyslipidemia, diabetes and coronary artery disease, showed a graded association with BMI, whilst smoking was most frequent in underweight patients. Correlation between BMI and WC was 0.748. As compared to normal weight patients' class III obese patients had highest mortality (HR 1.91; 95% CI: 1.650–2.202), following by class II obesity (HR 1.47, 95% CI: 1.33–1.62) and underweight patients (HR 1.47; 95% CI 1.1–1.8). Overweight patients had also slightly increased risk of dying as compared to normal weight patients (HR 1.12; 95% CI: 1.05–1.19). Similar U-shaped curve was observed for waist circumference. After adjustment for age, sex, smoking, diabetes, hypertension and LDL cholesterol highest mortality was observed for underweight patients (HR 1.88; 95% CI: 1.44–2.45), following by class III obesity (HR–1.51; 95% CI 1.31–1.75). Class I obesity and being overweight was associated with lower risk (HR 0.88; 95% CI: 0.82–0.94 and HR 0.86; 95% CI: 0.81–0.92, respectively). Conclusions No evidence of obesity paradox was present in unadjusted analysis. Lower mortality was observed in overweight and class I obese patients only after adjustment for risk factors and diseases typically associated with obesity. Results of multivariate analysis indicate that diseases that develop in obese patients are significantly associated with mortality more than obesity itself. Confounding may be responsible for discussed obesity paradox in literature therefore preventive measures to maintain BMI within normal range and prevent development of its complications should not be abolished. Funding Acknowledgement Type of funding sources: Private company. Main funding source(s): The present study was funded by an unrestricted educational grant from Valeant. As a supporter of the study, Valeant played no role in the study design, data analysis, data interpretation, or writing of the report. The present study was also supported by Silesian Analytical Laboratories (SLA, Katowice, Poland).

Time-fixed vs time-varying causal diagrams for immortal time bias.

We thank Shrier and Suissa1 for their interest in our editorial2 on causal diagrams3–7 for immortal time bias (ITB) and for their detailed analysis of our causal diagrams. They proposed a time-varying approach for these diagrams and concluded that the structure of ITB for both exposure misclassification and exclusion scenarios is selection (collider-stratification) bias. We chose to use simpler causal diagrams in our editorial in order to explain the key concepts without confusing the readers who might be less familiar with these diagrams, but we welcome the chance now to consider the more detailed analysis of ITB using time-varying diagrams. Here, we first explain why we believe that our time-fixed causal diagrams are correct. Then we will demonstrate that Shrier and Suissa’s presentation and interpretation of time-varying causal diagrams for ITB are incorrect, and also provide the correct time-varying diagrams that are actually equivalent to the time-fixed diagrams originally presented in our editorial.

Monte Carlo Simulation Approaches for Quantitative Bias Analysis: A Tutorial.

Quantitative bias analysis can be used to empirically assess how far study estimates are from the truth (i.e., an estimate that is free of bias). These methods can be used to explore the potential impact of confounding bias, selection bias (collider stratification bias), and information bias. Quantitative bias analysis includes methods that can be used to check the robustness of study findings to multiple types of bias and methods that use simulation studies to generate data and understand the hypothetical impact of specific types of bias in a simulated data set. In this article, we review 2 strategies for quantitative bias analysis: 1) traditional probabilistic quantitative bias analysis and 2) quantitative bias analysis with generated data. An important difference between the 2 strategies relates to the type of data (real vs. generated data) used in the analysis. Monte Carlo simulations are used in both approaches, but the simulation process is used for different purposes in each. For both approaches, we outline and describe the steps required to carry out the quantitative bias analysis and also present a bias-analysis tutorial demonstrating how both approaches can be applied in the context of an analysis for selection bias. Our goal is to highlight the utility of quantitative bias analysis for practicing epidemiologists and increase the use of these methods in the epidemiologic literature.

286Collider-stratification bias when estimating variable importance using Random Forests

Abstract Background Advances in causal inference have helped explain the longstanding birthweight and obesity paradoxes: selection bias due to conditioning on a collider variable i.e. collider-stratification bias (CSB). The lessons learned have critical implications for the interpretation of machine learning (ML), including decision trees and random forests (RFs), that implicitly condition on input variables. RFs are a popular approach for identifying important “predictors” from large data through variable importance, defined by the average decrease in prediction accuracy. While CSB has become a recognized concern when estimating exposure-outcome effects, knowledge of its impact on ML’s variable importance measures (VIMs) is limited. Applying the causal inference framework, we investigated the accuracy of RFs’ VIMs in data-mechanisms prone to CSB. Methods A Monte Carlo simulation study was conducted, with binary outcome and collider variables generated from logistic models. Two exposure variables stochastically determined the outcome and a collider variable, independent of the outcome. VIMs from RFs were compared to the known causal relevance of the input variables on the outcome. Results While variable importance of true exposure variables was not systematically affected by CSB, validity of VIMs can be affected, leading to erroneous selection of collider variables, causally independent of the outcome, as outcome predictors. Conclusions In presence of CSB, VIMs are not valid measures of the causal relevance of variables and may mislead selection of truly important factors that affect the outcome. Key messages ML must consider causal data-generating mechanisms otherwise it may lead to erroneous assessment of variable importance regarding outcome prediction.

Effectiveness of Aspirin in COPD: Biases in the Observational Studies

Several observational studies report decreased incidence of mortality and of exacerbations with aspirin use in patients with chronic obstructive pulmonary disease (COPD), with calls for a large randomized trial. Aspirin does have local and systemic pulmonary mechanisms of action that could make this drug beneficial in the treatment of COPD. However, the potential for biases in the observational studies has not been examined. We searched the literature for all observational studies reporting on the effect of aspirin in COPD patients on exacerbation and mortality. We reviewed the studies for the presence of time-related and other biases. We identified eight observational studies reporting an overall reduction in all-cause mortality or exacerbation with aspirin use of 21% (pooled rate ratio (RR) 0.79; 95% CI 0.71–0.86). We found two studies affected by immortal time bias (pooled RR 0.81; 95% CI 0.74–0.89), three studies affected by collider-stratification bias (pooled RR 0.66; 95% CI 0.55–0.79) and three that involved some exposure misclassification (pooled RR 0.85; 95% CI 0.78–0.92). Moreover, while adjusting for cardiovascular factors, six of the eight studies did not adjust for important markers of COPD severity and thus remain susceptible to confounding bias. In conclusion, all observational studies reporting on the effectiveness of aspirin on major outcomes of COPD are affected by biases known to exaggerate the effectiveness of a drug. As these studies cannot be used to support a beneficial effect for aspirin in COPD, it would be premature to consider a randomized trial to investigate this question until methodologically rigorous studies are available.