Reweighting UK Biobank corrects for pervasive selection bias due to volunteering.

BACKGROUND: Biobanks typically rely on volunteer-based sampling. This results in large samples (power) at the cost of representativeness (bias). The problem of volunteer bias is debated. Here, we (i) show that volunteering biases associations in UK Biobank (UKB) and (ii) estimate inverse probability (IP) weights that correct for volunteer bias in UKB. METHODS: Drawing on UK Census data, we constructed a subsample representative of UKB's target population, which consists of all individuals invited to participate. Based on demographic variables shared between the UK Census and UKB, we estimated IP weights (IPWs) for each UKB participant. We compared 21 weighted and unweighted bivariate associations between these demographic variables to assess volunteer bias. RESULTS: Volunteer bias in all associations, as naively estimated in UKB, was substantial-in some cases so severe that unweighted estimates had the opposite sign of the association in the target population. For example, older individuals in UKB reported being in better health, in contrast to evidence from the UK Census. Using IPWs in weighted regressions reduced 87% of volunteer bias on average. Volunteer-based sampling reduced the effective sample size of UKB substantially, to 32% of its original size. CONCLUSIONS: Estimates from large-scale biobanks may be misleading due to volunteer bias. We recommend IP weighting to correct for such bias. To aid in the construction of the next generation of biobanks, we provide suggestions on how to best ensure representativeness in a volunteer-based design. For UKB, IPWs have been made available.

Household income and health-related quality of life in children receiving treatment for acute myeloid leukemia: Potential impact of selection bias in health equity research.

OBJECTIVE: Examine the influence of household income on health-related quality of life (HRQOL) among children with newly diagnosed acute myeloid leukemia (AML). DESIGN: Secondary analysis of data prospectively collected from pediatric patients receiving treatment for AML at 14 hospitals across the United States. EXPOSURE: Household income was self-reported on a demographic survey. The examined mediators included the acuity of presentation and treatment toxicity. OUTCOME: Caregiver proxy reported assessment of patient HRQOL from the Peds QL 4.0 survey. RESULT: Children with AML (n = 131) and caregivers were prospectively enrolled to complete PedsQL assessments. HRQOL scores were better for patients in the lowest versus highest income category (mean ± SD: 76.0 ± 14 household income <$25,000 vs. 59.9 ± 17 income ≥$75,000; adjusted mean difference: 11.2, 95% CI: 2.2-20.2). Seven percent of enrolled patients presented with high acuity (ICU-level care in the first 72 h), and 16% had high toxicity (any ICU-level care); there were no identifiable differences by income, refuting mediating roles in the association between income and HRQOL. Enrolled patients were less likely to be Black/African American (9.9% vs. 22.2%), more likely to be privately insured (50.4% vs. 40.7%), and more likely to have been treated on a clinical trial (26.7% vs. 18.5%) compared to eligible unenrolled patients not enrolled. Evaluations of potential selection bias on the association between income and HRQOL suggested differences in HRQOL may be smaller than observed or even in the opposing direction. CONCLUSIONS: While primary analyses suggested lower household income was associated with superior HRQOL, differential participation may have biased these results. Future studies should partner with patients/families to identify strategies for equitable participation in clinical research.

Transcriptome analyses in juvenile yellow perch (Perca flavescens) exposed in vivo to clothianidin and chlorantraniliprole: Possible sampling bias.

The St. Lawrence River is an important North American waterway that is subject to anthropogenic pressures including intensive urbanization, and agricultural development. Pesticides are widely used for agricultural activities in fields surrounding the yellow perch (Perca flavescens) habitat in Lake St. Pierre (Quebec, Canada), a fluvial lake of the river where the perch population has collapsed. Clothianidin and chlorantraniliprole were two of the most detected insecticides in surface waters near perch spawning areas. The objectives of the present study were to evaluate the transcriptional and biochemical effects of these two pesticides on juvenile yellow perch exposed for 28d to environmental doses of each compound alone and in a mixture under laboratory/aquaria conditions. Hepatic mRNA-sequencing revealed an effect of chlorantraniliprole alone (37 genes) and combined with clothianidin (251 genes), but no effects of clothianidin alone were observed in perch. Dysregulated genes were mostly related to circadian rhythms and to Ca2+ signaling, the latter effect has been previously associated with chlorantraniliprole mode of action in insects. Moreover, chronic exposure to clothianidin increased the activity of acetylcholinesterase in the brain of exposed fish, suggesting a potential non-target effect of this insecticide. Further analyses of three clock genes by qRT-PCR suggested that part of the observed effects of chlorantraniliprole on the circadian gene regulation of juvenile perch could be the result of time-of-day of sacrifice. These results provide insight into biological effects of insecticides in juvenile perch and highlight the importance of considering the circadian rhythm in experimental design and results analyses.

Influence of School Type and Class Level on Mean Caries Experience in 12-Year-Olds in Serial Cross-Sectional National Oral Health Survey in Germany-Proposal to Adjust for Selection Bias.

The objective of this study is to analyse the effects of attended school type and class level on the reported caries experience (DMFT) obtained in the serial cross-sectional National Oral Health Study in Children in Germany (NOHSC) for the WHO reference group of 12-year-olds. METHODS: Caries data from the 2016 NOHSC were adjusted for each federal state on the basis of two additional large-scale datasets for school type and class level. RESULTS: Twelve-year-olds in all grades in Saxony-Anhalt (n = 96,842) exhibited significantly higher DMFT values than 12-year-olds in 6th grade (n = 76,456; +0.10 DMFT; ~14.2%, p < 0.001). Adjustments for school type had effects on DMFT on the level of federal states but almost balanced out on the national level (-0.01 DMFT; ~2%). Due to putatively similar structures of the federal states, the national mean DMFT for 12-year-olds in the latest NOHSC (2016; n = 55,002) was adjusted from 0.44 to 0.50 DMFT, correcting for selection bias. CONCLUSION: Selection bias in this NOHSC leads to an underestimation of caries levels by about 15%. Due to very low caries experience in children in Germany, these precise adjustments (+0.06 DMFT) have only a minor effect on interpretations of the national epidemiologic situation. Consequently, other national caries studies worldwide using the robust marker of DMFT should also adjust for systematic selection bias related to socio-economic background rather than increasing efforts in examination strategy.

Attention-deficit hyperactivity disorder symptoms and brain morphology: Addressing potential selection bias with inverse probability weighting.

The goal of this study was to examine what happens to established associations between attention deficit hyperactivity disorder (ADHD) symptoms and cortical surface and thickness regions once we apply inverse probability of censoring weighting (IPCW) to address potential selection bias. Moreover, we illustrate how different factors that predict participation contribute to potential selection bias. Participants were 9- to 11-year-old children from the Generation R study (N = 2707). Cortical area and thickness were measured with magnetic resonance imaging (MRI) and ADHD symptoms with the Child Behavior Checklist. We examined how associations between ADHD symptoms and brain morphology change when we weight our sample back to either follow-up (ages 9-11), baseline (cohort at birth), or eligible (population of Rotterdam at time of recruitment). Weights were derived using IPCW or raking and missing predictors of participation used to estimate weights were imputed. Weighting analyses to baseline and eligible increased beta coefficients for the middle temporal gyrus surface area, as well as fusiform gyrus cortical thickness. Alternatively, the beta coefficient for the rostral anterior cingulate decreased. Removing one group of variables used for estimating weights resulted in the weighted regression coefficient moving closer to the unweighted regression coefficient. In addition, we found considerably different beta coefficients for most surface area regions and all thickness measures when we did not impute missing covariate data. Our findings highlight the importance of using inverse probability weighting (IPW) in the neuroimaging field, especially in the context of mental health-related research. We found that including all variables related to exposure-outcome in the IPW model and combining IPW with multiple imputations can help reduce bias. We encourage future psychiatric neuroimaging studies to define their target population, collect information on eligible but not included participants and use inverse probability of censoring weighting (IPCW) to reduce selection bias.

Bias of the additive hazard model in the presence of causal effect heterogeneity.

Hazard ratios are prone to selection bias, compromising their use as causal estimands. On the other hand, if Aalen's additive hazard model applies, the hazard difference has been shown to remain unaffected by the selection of frailty factors over time. Then, in the absence of confounding, observed hazard differences are equal in expectation to the causal hazard differences. However, in the presence of effect (on the hazard) heterogeneity, the observed hazard difference is also affected by selection of survivors. In this work, we formalize how the observed hazard difference (from a randomized controlled trial) evolves by selecting favourable levels of effect modifiers in the exposed group and thus deviates from the causal effect of interest. Such selection may result in a non-linear integrated hazard difference curve even when the individual causal effects are time-invariant. Therefore, a homogeneous time-varying causal additive effect on the hazard cannot be distinguished from a time-invariant but heterogeneous causal effect. We illustrate this causal issue by studying the effect of chemotherapy on the survival time of patients suffering from carcinoma of the oropharynx using data from a clinical trial. The hazard difference can thus not be used as an appropriate measure of the causal effect without making untestable assumptions.

Unmasking bias in artificial intelligence: a systematic review of bias detection and mitigation strategies in electronic health record-based models.

OBJECTIVES: Leveraging artificial intelligence (AI) in conjunction with electronic health records (EHRs) holds transformative potential to improve healthcare. However, addressing bias in AI, which risks worsening healthcare disparities, cannot be overlooked. This study reviews methods to handle various biases in AI models developed using EHR data. MATERIALS AND METHODS: We conducted a systematic review following the Preferred Reporting Items for Systematic Reviews and Meta-analyses guidelines, analyzing articles from PubMed, Web of Science, and IEEE published between January 01, 2010 and December 17, 2023. The review identified key biases, outlined strategies for detecting and mitigating bias throughout the AI model development, and analyzed metrics for bias assessment. RESULTS: Of the 450 articles retrieved, 20 met our criteria, revealing 6 major bias types: algorithmic, confounding, implicit, measurement, selection, and temporal. The AI models were primarily developed for predictive tasks, yet none have been deployed in real-world healthcare settings. Five studies concentrated on the detection of implicit and algorithmic biases employing fairness metrics like statistical parity, equal opportunity, and predictive equity. Fifteen studies proposed strategies for mitigating biases, especially targeting implicit and selection biases. These strategies, evaluated through both performance and fairness metrics, predominantly involved data collection and preprocessing techniques like resampling and reweighting. DISCUSSION: This review highlights evolving strategies to mitigate bias in EHR-based AI models, emphasizing the urgent need for both standardized and detailed reporting of the methodologies and systematic real-world testing and evaluation. Such measures are essential for gauging models' practical impact and fostering ethical AI that ensures fairness and equity in healthcare.

Towards a Clearer Causal Question Underlying the Association Between Cancer and Dementia.

BACKGROUND: Several observational studies have described an inverse association between cancer diagnosis and subsequent dementia risk. Multiple biologic mechanisms and potential biases have been proposed in attempts to explain this association. One proposed explanation is the opposite expression of Pin1 in cancer and dementia, and we use this explanation and potential drug target to illustrate the required assumptions and potential sources of bias for inferring an effect of Pin1 on dementia risk from analyses measuring cancer diagnosis as a proxy for Pin1 expression. METHODS: We used data from the Rotterdam Study, a population-based cohort. We estimate the association between cancer diagnosis (as a proxy for Pin1) and subsequent dementia diagnosis using two different proxy methods and with confounding and censoring for death addressed with inverse probability weights. We estimate and compare the complements of a weighted Kaplan-Meier survival estimator at 20 years of follow-up. RESULTS: Out of 3634 participants, 899 (25%) were diagnosed with cancer, of whom 53 (6%) had dementia, and 567 (63%) died. Among those without cancer, 15% (411) were diagnosed with dementia, and 667 (24%) died over follow-up. Depending on the confounding and selection bias control, and the way in which cancer was used as a time-varying proxy exposure, the risk ratio for dementia diagnosis ranged from 0.71 (95% confidence interval [CI] = 0.49, 0.95) to 1.1 (95% CI = 0.79, 1.3). CONCLUSION: Being explicit about the underlying mechanism of interest is key to maximizing what we can learn from this cancer-dementia association given available or readily collected data, and to defining, detecting, and preventing potential biases.

A class of computational methods to reduce selection bias when designing Phase 3 clinical trials.

When designing confirmatory Phase 3 studies, one usually evaluates one or more efficacious and safe treatment option(s) based on data from previous studies. However, several retrospective research articles reported the phenomenon of "diminished treatment effect in Phase 3" based on many case studies. Even under basic assumptions, it was shown that the commonly used estimator could substantially overestimate the efficacy of selected group(s). As alternatives, we propose a class of computational methods to reduce estimation bias and mean squared error with a broader scope of multiple treatment groups and flexibility to accommodate summary results by group as input. Based on simulation studies and a real data example, we provide practical implementation guidance for this class of methods under different scenarios. For more complicated problems, our framework can serve as a starting point with additional layers built in. Proposed methods can also be widely applied to other selection problems.

Bias correction models for electronic health records data in the presence of non-random sampling.

Electronic health records (EHRs) contain rich clinical information for millions of patients and are increasingly used for public health research. However, non-random inclusion of subjects in EHRs can result in selection bias, with factors such as demographics, socioeconomic status, healthcare referral patterns, and underlying health status playing a role. While this issue has been well documented, little work has been done to develop or apply bias-correction methods, often due to the fact that most of these factors are unavailable in EHRs. To address this gap, we propose a series of Heckman type bias correction methods by incorporating social determinants of health selection covariates to model the EHR non-random sampling probability. Through simulations under various settings, we demonstrate the effectiveness of our proposed method in correcting biases in both the association coefficient and the outcome mean. Our method augments the utility of EHRs for public health inferences, as we show by estimating the prevalence of cardiovascular disease and its correlation with risk factors in the New York City network of EHRs.

Strong positive selection biases identity-by-descent-based inferences of recent demography and population structure in Plasmodium falciparum.

Malaria genomic surveillance often estimates parasite genetic relatedness using metrics such as Identity-By-Decent (IBD), yet strong positive selection stemming from antimalarial drug resistance or other interventions may bias IBD-based estimates. In this study, we use simulations, a true IBD inference algorithm, and empirical data sets from different malaria transmission settings to investigate the extent of this bias and explore potential correction strategies. We analyze whole genome sequence data generated from 640 new and 3089 publicly available Plasmodium falciparum clinical isolates. We demonstrate that positive selection distorts IBD distributions, leading to underestimated effective population size and blurred population structure. Additionally, we discover that the removal of IBD peak regions partially restores the accuracy of IBD-based inferences, with this effect contingent on the population's background genetic relatedness and extent of inbreeding. Consequently, we advocate for selection correction for parasite populations undergoing strong, recent positive selection, particularly in high malaria transmission settings.

The built-in selection bias of hazard ratios formalized using structural causal models.

It is known that the hazard ratio lacks a useful causal interpretation. Even for data from a randomized controlled trial, the hazard ratio suffers from so-called built-in selection bias as, over time, the individuals at risk among the exposed and unexposed are no longer exchangeable. In this paper, we formalize how the expectation of the observed hazard ratio evolves and deviates from the causal effect of interest in the presence of heterogeneity of the hazard rate of unexposed individuals (frailty) and heterogeneity in effect (individual modification). For the case of effect heterogeneity, we define the causal hazard ratio. We show that the expected observed hazard ratio equals the ratio of expectations of the latent variables (frailty and modifier) conditionally on survival in the world with and without exposure, respectively. Examples with gamma, inverse Gaussian and compound Poisson distributed frailty and categorical (harming, beneficial or neutral) distributed effect modifiers are presented for illustration. This set of examples shows that an observed hazard ratio with a particular value can arise for all values of the causal hazard ratio. Therefore, the hazard ratio cannot be used as a measure of the causal effect without making untestable assumptions, stressing the importance of using more appropriate estimands, such as contrasts of the survival probabilities.

Footprint of publication selection bias on meta-analyses in medicine, environmental sciences, psychology, and economics.

Publication selection bias undermines the systematic accumulation of evidence. To assess the extent of this problem, we survey over 68,000 meta-analyses containing over 700,000 effect size estimates from medicine (67,386/597,699), environmental sciences (199/12,707), psychology (605/23,563), and economics (327/91,421). Our results indicate that meta-analyses in economics are the most severely contaminated by publication selection bias, closely followed by meta-analyses in environmental sciences and psychology, whereas meta-analyses in medicine are contaminated the least. After adjusting for publication selection bias, the median probability of the presence of an effect decreased from 99.9% to 29.7% in economics, from 98.9% to 55.7% in psychology, from 99.8% to 70.7% in environmental sciences, and from 38.0% to 29.7% in medicine. The median absolute effect sizes (in terms of standardized mean differences) decreased from d = 0.20 to d = 0.07 in economics, from d = 0.37 to d = 0.26 in psychology, from d = 0.62 to d = 0.43 in environmental sciences, and from d = 0.24 to d = 0.13 in medicine.

Biased effects of pre-diagnostic physical activity on breast cancer survival: Systematic review and meta-analysis.

BACKGROUND: Pre-diagnostic physical activity is reported to improve survival for women with breast cancer. However, studies of pre-diagnostic exposures and cancer survival are susceptible to bias, made clear when applying a target trial framework. We investigated the impact of selection bias, immortal time bias, confounding and bias due to inappropriate adjustment for post-exposure variables in a systematic review and meta-analysis of pre-diagnostic physical activity and survival after breast cancer. METHODS: Medline, Embase and Emcare were searched from inception to November 2021 for studies examining pre-diagnostic physical activity and overall or breast cancer-specific survival for women with breast cancer. Random-effects meta-analysis was used to estimate pooled hazard ratios (HRs) and 95% confidence intervals (CIs) comparing highest versus lowest pre-diagnostic physical activity. Subgroup meta-analyses were used to compare HRs of studies with and without different biases. ROBINS-E was used to assess risk of bias. RESULTS: We included 22 studies. Women with highest versus lowest pre-diagnostic physical activity had higher overall and breast cancer-specific survival across most analyses. The overall risk of bias was high. We observed marked differences in estimated HRs between studies that did and did not adjust for post-exposure variables or have immortal time bias. All studies were at risk of selection bias due to participants becoming eligible for study when they have survived to post-exposure events (e.g., breast cancer diagnosis). Insufficient studies were available to investigate confounding. CONCLUSION: Biases can substantially change effect estimates. Due to misalignment of treatment assignment (before diagnosis), eligibility (survival to post-exposure events) and start of follow-up, bias is difficult to avoid. It is difficult to lend a causal interpretation to effect estimates from studies of pre-diagnostic physical activity and survival after cancer. Biased effect estimates that are difficult to interpret may be less useful for clinical or public health policy applications.

Revising model for end-stage liver disease from calendar-time cross-sections with correction for selection bias.

BACKGROUND: Eurotransplant liver transplant candidates are prioritized by Model for End-stage Liver Disease (MELD), a 90-day waitlist survival risk score based on the INR, creatinine and bilirubin. Several studies revised the original MELD score, UNOS-MELD, with transplant candidate data by modelling 90-day waitlist mortality from waitlist registration, censoring patients at delisting or transplantation. This approach ignores biomarkers reported after registration, and ignores informative censoring by transplantation and delisting. METHODS: We study how MELD revision is affected by revision from calendar-time cross-sections and correction for informative censoring with inverse probability censoring weighting (IPCW). For this, we revised UNOS-MELD on patients with chronic liver cirrhosis on the Eurotransplant waitlist between 2007 and 2019 (n = 13,274) with Cox models with as endpoints 90-day survival (a) from registration and (b) from weekly drawn calendar-time cross-sections. We refer to the revised score from cross-section with IPCW as DynReMELD, and compare DynReMELD to UNOS-MELD and ReMELD, a prior revision of UNOS-MELD for Eurotransplant, in geographical validation. RESULTS: Revising MELD from calendar-time cross-sections leads to significantly different MELD coefficients. IPCW increases estimates of absolute 90-day waitlist mortality risks by approximately 10 percentage points. DynReMELD has improved discrimination over UNOS-MELD (delta c-index: 0.0040, p < 0.001) and ReMELD (delta c-index: 0.0015, p < 0.01), with differences comparable in magnitude to the addition of an extra biomarker to MELD (delta c-index: ± 0.0030). CONCLUSION: Correcting for selection bias by transplantation/delisting does not improve discrimination of revised MELD scores, but substantially increases estimated absolute 90-day mortality risks. Revision from cross-section uses waitlist data more efficiently, and improves discrimination compared to revision of MELD exclusively based on information available at listing.

Occupational solar exposure and basal cell carcinoma. A review of the epidemiologic literature with meta-analysis focusing on particular methodological aspects.

BACKGROUND: Numerous epidemiologic studies and a few systematic reviews have investigated the association between occupational solar exposure and basal cell carcinoma (BCC). However, previous reviews have several deficits with regard to included and excluded studies/risk estimates and the assessment of risk of selection bias (RoSB). Our aim was to review epidemiologic studies with a focus on these deficits and to use meta-(regression) analyses to summarize risk estimates. METHODS: We systematically searched PubMed (including MEDLINE) and Embase for epidemiologic studies. Study evaluation considered four main aspects of risk of bias assessments, i.e. Selection of subjects (selection bias); Exposure variables; Outcome variables; Data analysis. RESULTS: Of 56 identified references, 32 were used for meta-(regression) analyses. The overall pooled risk estimate for BCC comparing high/present vs. low/absent occupational solar exposure was 1.20 (95% CI 1.02-1.43); among studies without major deficits regarding data analysis, it was 1.10 (95% CI 0.91-1.33). Studies with low and high RoSB had pooled risk estimates of 0.83 (95% CI 0.73-0.93) and 1.95 (95% CI 1.42-2.67), respectively. The definitions of exposure and outcome variables were not correlated with study risk estimates. Studies with low RoSB in populations with the same latitude or lower than Germany had a pooled risk estimate of 1.01 (95% CI 0.88-1.15). CONCLUSION: Due to the different associations between occupational solar exposure and BCC among studies with low and high RoSB, we reason that the current epidemiologic evidence base does not permit the conclusion that regular outdoor workers have an increased risk of BCC.

Effector-independent Representations Guide Sequential Target Selection Biases in Action.

Previous work shows that automatic attention biases toward recently selected target features transfer across action and perception and even across different effectors such as the eyes and hands on a trial-by-trial basis. Although these findings suggest a common neural representation of selection history across effectors, the extent to which information about recently selected target features is encoded in overlapping versus distinct brain regions is unknown. Using fMRI and a priming of pop-out task where participants selected unpredictable, uniquely colored targets among homogeneous distractors via reach or saccade, we show that color priming is driven by shared, effector-independent underlying representations of recent selection history. Consistent with previous work, we found that the intraparietal sulcus (IPS) was commonly activated on trials where target colors were switched relative to those where the colors were repeated; however, the dorsal anterior insula exhibited effector-specific activation related to color priming. Via multivoxel cross-classification analyses, we further demonstrate that fine-grained patterns of activity in both IPS and the medial temporal lobe encode information about selection history in an effector-independent manner, such that ROI-specific models trained on activity patterns during reach selection could predict whether a color was repeated or switched on the current trial during saccade selection and vice versa. Remarkably, model generalization performance in IPS and medial temporal lobe also tracked individual differences in behavioral priming sensitivity across both types of action. These results represent a first step to clarify the neural substrates of experience-driven selection biases in contexts that require the coordination of multiple actions.

Steady-state statistical properties and implementation of randomization designs with maximum tolerated imbalance restriction for two-arm equal allocation clinical trials.

In recent decades, several randomization designs have been proposed in the literature as better alternatives to the traditional permuted block design (PBD), providing higher allocation randomness under the same restriction of the maximum tolerated imbalance (MTI). However, PBD remains the most frequently used method for randomizing subjects in clinical trials. This status quo may reflect an inadequate awareness and appreciation of the statistical properties of these randomization designs, and a lack of simple methods for their implementation. This manuscript presents the analytic results of statistical properties for five randomization designs with MTI restriction based on their steady-state probabilities of the treatment imbalance Markov chain and compares them to those of the PBD. A unified framework for randomization sequence generation and real-time on-demand treatment assignment is proposed for the straightforward implementation of randomization algorithms with explicit formulas of conditional allocation probabilities. Topics associated with the evaluation, selection, and implementation of randomization designs are discussed. It is concluded that for two-arm equal allocation trials, several randomization designs offer stronger protection against selection bias than the PBD does, and their implementation is not necessarily more difficult than the implementation of the PBD.