Department of Biostatistics Seminar/Workshop Series

Biostatistics Student Research Forum:

Improving statistical estimation in observational data by using prediction models derived from partially validated data

Mark Giganti, PhD Candidate, Departmetn of Biostatistics

Vanderbilt University, School of Medicine

Observational data, collected in many health care settings throughout the world, provides a cost-effective resource for medical research. This data, however, is often of uncertain quality and may have incorrectly entered values. Considering error-prone data has the potential to significantly bias results, it is important to have an effective strategy for identifying such errors and resolving them.

Many different strategies exist for addressing observational study data quality. Among them are: ignoring the problem, reviewing the critical data fields in every patient and re-entering incorrectly entered values, or using a subset of reviewed patient charts to make a decision to either review all files (if the error rate is high enough) or ignoring the problem (if the error rate is low enough). Such data auditing and validation procedures, however, are resource-intensive.

In this talk, I explore the feasibility of validating a sub-sample of records and using statistical models of the error patterns in that subset to predict errors in the remaining unvalidated data. The talk will briefly review key concepts in measurement error, missing data, and prediction models. The strengths and shortcomings of these approaches will be contrasted to motivate a proposed prediction strategy that is derived from a (small) validation subset. Preliminary findings based on data from the Vanderbilt Comprehensive Care Clinic - where comprehensive chart reviews were performed to validate data for key variables - will be presented and discussed. These data have been fully validated, providing a unique opportunity to fully evaluate the performance and feasibility of our proposed approach.