Electronic health records (EHRs) are increasingly recognized as a cost-effective resource for patient recruitment in clinical research. However, how to optimally select a cohort from millions of individuals to answer a scientific question of interest remains unclear. Consider a study to estimate the mean or mean difference of an expensive outcome. Inexpensive auxiliary covariates predictive of the outcome may often be available in patients\' health records, presenting an opportunity to recruit patients selectively, which may improve efficiency in downstream analyses. In this paper we propose a two-phase sampling design that leverages available information on auxiliary covariates in EHR data. A key challenge in using EHR data for multiphase sampling is the potential selection bias, because EHR data are not necessarily representative of the target population. Extending existing literature on two-phase sampling design, we derive an optimal two-phase sampling method that improves efficiency over random sampling while accounting for the potential selection bias in EHR data. We demonstrate the efficiency gain from our sampling design via simulation studies and an application evaluating the prevalence of hypertension among U.S. adults leveraging data from the Michigan Genomics Initiative, a longitudinal biorepository in Michigan Medicine.

BACKGROUND: Disease latency is defined as the time from disease initiation to disease diagnosis. Disease latency bias (DLB) can arise in epidemiological studies that examine latent outcomes, since the exact timing of the disease inception is unknown and might occur before exposure initiation, potentially leading to bias. Although DLB can affect epidemiological studies that examine different types of chronic disease (e.g. Alzheimer\'s disease, cancer etc), the manner by which DLB can introduce bias into these studies has not been previously elucidated. Information on the specific types of bias, and their structure, that can arise secondary to DLB is critical for researchers, to enable better understanding and control for DLB.
METHODS: Here we describe four scenarios by which DLB can introduce bias (through different structures) into epidemiological studies that address latent outcomes, using directed acyclic graphs (DAGs). We also discuss potential strategies to better understand, examine and control for DLB in these studies.
CONCLUSIONS: Using causal diagrams, we show that disease latency bias can affect results of epidemiological studies through: (i) unmeasured confounding; (ii) reverse causality; (iii) selection bias; (iv) bias through a mediator.
CONCLUSIONS: Disease latency bias is an important bias that can affect a number of epidemiological studies that address latent outcomes. Causal diagrams can assist researchers better identify and control for this bias.

Selection bias has long been central in methodological discussions across epidemiology and other fields. In epidemiology, the concept of selection bias has been continually evolving over time. In this issue of the Journal, Mathur and Shpitser (Am J Epidemiol. XXXX;XXX(XX):XXXX-XXXX) present simple graphical rules for using a Single World Intervention Graph (SWIG) to assess the presence of selection bias when estimating treatment effects in both the general population and a selected sample. Notably, the authors examine the setting in which the treatment affects selection, an issue not well-addressed in the existing literature on selection bias. To place the work by Mathur and Shpitser in context, we review the evolution of the concept of selection bias in epidemiology, with a primary focus on the developments in the last 20-30 years since the introduction of causal directed acyclic graphs (DAGs) to epidemiologic research.

暂无摘要。

Our study aimed to establish the risk of selection bias in randomized controlled trials (RCT) that were overall rated as having \"low bias\" risk according to Cochrane\'s Risk of Bias, version 2 (RoB 2) tool. A systematic literature search of current systematic reviews of RCTs was conducted. From the identified reviews, RCTs with overall \"high bias\" and \"low bias\" RoB 2 risk ratings were extracted. All RCTs were statistically tested for selection bias risk. From the test results, true positive, true negative, false positive, or false negative ratings were established, and the false omission rate (FOR) with a 95% confidence interval (CI) was computed. Subgroup analysis was conducted by computing the negative likelihood ratio (-LR) concerning RoB 2 domain 1 ratings: bias arising from the randomization process. A total of 1070 published RCTs (median publication year: 2018; interquartile range: 2013-2020) were identified and tested. We found that 7.61% of all \"low bias\" (RoB 2)-rated RCTs were of high selection bias risk (FOR 7.61%; 95% CI: 6.31%-9.14%) and that the likelihood for high selection bias risk in \"low bias\" (RoB 2 domain 1)-rated RCTs was 6% higher than that for low selection bias risk (-LR: 1.06; 95% CI: 0.98-1.15). These findings raise issues about the validity of \"low bias\" risk ratings using Cochrane\'s RoB 2 tool as well as about the validity of some of the results from recently published RCTs. Our results also suggest that the likelihood of a \"low bias\" risk-rated body of clinical evidence being actually bias-free is low, and that generalization based on a limited, pre-specified set of appraisal criteria may not justify a high level of confidence that such evidence reflects the true treatment effect.

Observational studies are rarely representative of their target population because there are known and unknown factors that affect an individual\'s choice to participate (the selection mechanism). Selection can cause bias in a given analysis if the outcome is related to selection (conditional on the other variables in the model). Detecting and adjusting for selection bias in practice typically requires access to data on nonselected individuals. Here, we propose methods to detect selection bias in genetic studies by comparing correlations among genetic variants in the selected sample to those expected under no selection. We examine the use of four hypothesis tests to identify induced associations between genetic variants in the selected sample. We evaluate these approaches in Monte Carlo simulations. Finally, we use these approaches in an applied example using data from the UK Biobank (UKBB). The proposed tests suggested an association between alcohol consumption and selection into UKBB. Hence, UKBB analyses with alcohol consumption as the exposure or outcome may be biased by this selection.

BACKGROUND: Randomized controlled trials are the gold standard for determining treatment efficacy in medicine. To deter harmful practices such as p-hacking and hypothesizing after the results are known, any analysis of subgroups and secondary outcomes must be documented and pre-specified. However, they can still introduce bias (and routinely do) if they are not treated with the same consideration as the primary analysis.
METHODS: We describe several sources of bias that affect subgroup and secondary outcome analyses using published randomized trials and causal directed acyclic graphs (DAGs).
RESULTS: We use the RECOVERY and START trials to elucidate sources of bias in analyses of subgroups and secondary outcomes. Chance imbalance can occur if the distribution of prognostic variables is not sought for any given subgroup analysis as for the main analysis. This differential distribution of prognostic variables can also occur in analyses of secondary outcomes. Selection bias can occur if the subgroup variable is causally related to staying in the trial. Given loss to follow up is not normally addressed in subgroups, attrition bias can pass unnoticed in these cases. In every case, the solution is to take the same considerations for these analyses as we do for primary analyses.
CONCLUSIONS: Approval of treatments and clinical decisions can occur based on results from subgroup or secondary outcome analyses. Thus, it is important to give them the same treatment as primary analyses to avoid preventable biases.

Selection bias is a common concern in epidemiologic studies. In the literature, selection bias is often viewed as a missing data problem. Popular approaches to adjust for bias due to missing data, such as inverse probability weighting, rely on the assumption that data are missing at random and can yield biased results if this assumption is violated. In observational studies with outcome data missing not at random, Heckman\'s sample selection model can be used to adjust for bias due to missing data. In this paper, we review Heckman\'s method and a similar approach proposed by Tchetgen Tchetgen and Wirth (2017). We then discuss how to apply these methods to Mendelian randomization analyses using individual-level data, with missing data for either the exposure or outcome or both. We explore whether genetic variants associated with participation can be used as instruments for selection. We then describe how to obtain missingness-adjusted Wald ratio, two-stage least squares and inverse variance weighted estimates. The two methods are evaluated and compared in simulations, with results suggesting that they can both mitigate selection bias but may yield parameter estimates with large standard errors in some settings. In an illustrative real-data application, we investigate the effects of body mass index on smoking using data from the Avon Longitudinal Study of Parents and Children.

Consent bias is a type of selection bias in biomedical research where those consenting to the research differ systematically from those not consenting. It is particularly relevant in precision medicine research because the complexity of these studies prevents certain subgroups from understanding, trusting, and consenting to the research. Because consent bias distorts research findings and causes inequitable distribution of research benefits, scholars propose two types of schemes to reduce consent bias: reforming existing consent models and removing the consent requirement altogether. This study explores the possibility of waiving consent in observational studies using existing data, because they involve fewer risks to participants than clinical trials if privacy safeguards are strengthened. It suggests that data protection mechanisms such as security enhancement and data protection impact assessment should be conducted to protect data privacy of participants in observational studies without consent.

BACKGROUND: In population-based research, pregnancy may be a repeated event. Despite published guidance on how to address repeated pregnancies to the same individual, a variety of approaches are observed in perinatal epidemiological studies. While some of these approaches are supported by the chosen research question, others are consequences of constraints inherent to a given dataset (eg, missing parity information). These decisions determine how appropriately a given research question can be answered and overall generalizability.
OBJECTIVE: To compare common cohort selection and analytic approaches used for perinatal epidemiological research by assessing the prevalence of two perinatal outcomes and their association with a clinical and a social independent variable.
METHODS: Using vital records linked to maternal hospital discharge records for singleton births, we created four cohorts: (1) all-births (2) randomly selected one birth per individual (3) first-observed birth per individual (4) primiparous-births (parity 1). Sampling of births was not conditional on cluster (ie, we did not sample all births by a given mother, but rather sampled individual births). Study outcomes were severe maternal morbidity (SMM) and preeclampsia/eclampsia, and the independent variables were self-reported race/ethnicity (as a social factor) and systemic lupus erythematosus. Comparing the four cohorts, we assessed the distribution of maternal characteristics, the prevalence of outcomes, overall and stratified by parity, and risk ratios (RR) for the associations of outcomes with independent variables. Among all-births, we then compared RR from three analytic strategies: with standard inference that assumes independently sampled births to the same mother in the model, with cluster-robust inference, and adjusting for parity.
RESULTS: We observed minor differences in the population characteristics between the all-birth (N=2736,693), random-selection, and first-observed birth cohorts (both N=2284,660), with more substantial differences between these cohorts and the primiparous-births cohort (N=1054,684). Outcome prevalence was consistently lowest among all-births and highest among primiparous-births (eg, SMM 18.9 per 1000 births among primiparous-births vs 16.6 per 1000 births among all-births). When stratified by parity, outcome prevalence was always the lowest in births of parity 2 and highest among births of parity 1 for both outcomes. RR differed for study outcomes across all four cohorts, with the most pronounced differences between the primiparous-birth cohort and other cohorts. Among all-births, robust inference minimally impacted the confidence bounds of estimates, compared to the standard inference, that is, crude estimates (eg, lupus-SMM association: 4.01, 95% confidence intervals [CI] 3.54-4.55 vs 4.01, 95% CI 3.53-4.56 for crude estimate), while adjusting for parity slightly shifted estimates, toward the null for SMM and away from the null for preeclampsia/eclampsia.
CONCLUSIONS: Researchers should consider the alignment between the methods they use, their sampling strategy, and their research question. This could include refining the research question to better match inference possible for available data, considering alternative data sources, and appropriately noting data limitations and resulting bias, as well as the generalizability of findings. If parity is an established effect modifier, stratified results should be presented.