Dataset Information

Bias reduction and inference for electronic health record data under selection and phenotype misclassification: three case studies.

ABSTRACT: Electronic Health Records (EHR) are not designed for population-based research, but they provide access to longitudinal health information for many individuals. Many statistical methods have been proposed to account for selection bias, missing data, phenotyping errors, or other problems that arise in EHR data analysis. However, addressing multiple sources of bias simultaneously is challenging. Recently, we developed a methodological framework (R package, SAMBA ) for jointly handling both selection bias and phenotype misclassification in the EHR setting that leverages external data sources. These methods assume factors related to selection and misclassification are fully observed, but these factors may be poorly understood and partially observed in practice. As a follow-up to the methodological work, we explore how these methods perform for three real-world case studies. In all three examples, we use individual patient-level data collected through the University of Michigan Health System and various external population-based data sources. In case study (a), we explore the impact of these methods on estimated associations between gender and cancer diagnosis. In case study (b), we compare corrected associations between previously identified genetic loci and age-related macular degeneration with gold standard external estimates. In case study (c), we evaluate these methods for modeling the association of COVID-19 outcomes and potential risk factors. These case studies illustrate how to utilize diverse auxiliary information to achieve less biased inference in EHR-based research.

SUBMITTER: Beesley LJ

PROVIDER: S-EPMC7781342 | biostudies-literature | 2020 Dec

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Bias reduction and inference for electronic health record data under selection and phenotype misclassification: three case studies.

Beesley Lauren J LJ Mukherjee Bhramar B

medRxiv : the preprint server for health sciences 20201223

Electronic Health Records (EHR) are not designed for population-based research, but they provide access to longitudinal health information for many individuals. Many statistical methods have been proposed to account for selection bias, missing data, phenotyping errors, or other problems that arise in EHR data analysis. However, addressing multiple sources of bias simultaneously is challenging. Recently, we developed a methodological framework (R package, <i>SAMBA</i> ) for jointly handling both ...[more]

PMID: 33398299

Similar Datasets

Project description:ImportanceThe Centers for Medicare and Medicaid Services (CMS) use point estimates of 30-day risk-standardized readmission rates (RSRRs) to compare hospitals under the Hospital Readmissions Reduction Program (HRRP). An important characteristic of this measure is that it is a point estimate with a margin of error, which may affect the CMS's ability to accurately evaluate and distinguish hospital performance in the program.ObjectiveTo determine the number and percentage of hospitals with a penalty status misclassified under the HRRP.Design, setting, and participantsThis cross-sectional study used the bayesian deconvolution method to estimate the rate of penalty status misclassification for hospitals participating in the HRRP in fiscal year 2019, using data from the CMS Hospital Compare website that were collected between July 1, 2014, and June 30, 2017. Beneficiaries of Medicare fee-for-service coverage who were 65 years or older and hospitalized with acute myocardial infarction, heart failure, or pneumonia in hospitals with 25 or more discharges per condition were included in the data set. Data analysis occurred from November 2019 to July 2020.ExposuresNone.Main outcomes and measuresThe rate of penalty status misclassification for acute myocardial infarction, heart failure, or pneumonia under the HRRP.ResultsThe study included 1633, 2626, and 2705 hospitals for acute myocardial infarction, heart failure, and pneumonia, respectively, that participated in the HRRP in fiscal year 2019. The percentages of hospitals that should have been penalized, but were not, were 20.9% (95% CI, 16.0%-25.8%) for acute myocardial infarction, 13.5% (95% CI, 9.8%-17.2%) for heart failure, and 13.2% (95% CI, 10.3%-16.1%) for pneumonia. In contrast, the percentages of hospitals that were incorrectly penalized by the HRRP were 10.1% (95% CI, 5.8%-14.4%) for acute myocardial infarction, 10.9% (95% CI, 7.2%-14.6%) for heart failure, and 12.3% (95% CI, 9.9%-14.6%) for pneumonia.Conclusions and relevanceThe margin of error associated with the 30-day RSRRs resulted in the misclassification of condition-specific penalty status for up to 31% of hospitals. These findings suggest that the hospital-level 30-day RSRR measure may not reliably distinguish hospital performance in the HRRP. This has important implications for CMS value-based programs that use risk-standardized outcomes to evaluate and compare hospital performance.

Project description:Using imperfect tests may lead to biased estimates of disease frequency and measures of association. Many studies have looked into the effect of misclassification on statistical inferences. These evaluations were either within a cross-sectional study framework, assessing biased prevalence, or for cohort study designs, evaluating biased incidence rate or risk ratio estimates based on misclassification at one of the two time-points (initial assessment or follow-up). However, both observations at risk and incident cases can be wrongly identified in longitudinal studies, leading to selection and misclassification biases, respectively. The objective of this paper was to evaluate the relative impact of selection and misclassification biases resulting from misclassification, together, on measures of incidence and risk ratio. To investigate impact on measure of disease frequency, data sets from a hypothetical cohort study with two samples collected one month apart were simulated and analyzed based on specific test and disease characteristics, with no elimination of disease during the sampling interval or clustering of observations. Direction and magnitude of bias due to selection, misclassification, and total bias was assessed for diagnostic test sensitivity and specificity ranging from 0.7 to 1.0 and 0.8 to 1.0, respectively, and for specific disease contexts, i.e., disease prevalences of 5 and 20%, and disease incidences of 0.01, 0.05, and 0.1 cases/animal-month. A hypothetical exposure with known strength of association was also generated. A total of 1,000 cohort studies of 1,000 observations each were simulated for these six disease contexts where the same diagnostic test was used to identify observations at risk at beginning of the cohort and incident cases at its end. Our results indicated that the departure of the estimates of disease incidence and risk ratio from their true value were mainly a function of test specificity, and disease prevalence and incidence. The combination of the two biases, at baseline and follow-up, revealed the importance of a good to excellent specificity relative to sensitivity for the diagnostic test. Small divergence from perfect specificity extended quickly to disease incidence over-estimation as true prevalence increased and true incidence decreased. A highly sensitive test to exclude diseased subjects at baseline was of less importance to minimize bias than using a highly specific one at baseline. Near perfect diagnostic test attributes were even more important to obtain a measure of association close to the true risk ratio, according to specific disease characteristics, especially its prevalence. Low prevalent and high incident disease lead to minimal bias if disease is diagnosed with high sensitivity and close to perfect specificity at baseline and follow-up. For more prevalent diseases we observed large risk ratio biases towards the null value, even with near perfect diagnosis.

Dataset Information

Bias reduction and inference for electronic health record data under selection and phenotype misclassification: three case studies.

Publications

Bias reduction and inference for electronic health record data under selection and phenotype misclassification: three case studies.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure