Adjusting for Measurement Error in the Performance Evaluation of Diagnostic Medical Tests

• Activity Number: 25
• Type: Topic Contributed
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Gene Pennello*+
• Companies: Food and Drug Administration
• Keywords: nondifferential measurement error ; conditional independence ; misclassification ; structural model ; latent variable ; companion diagnostic

Abstract

Diagnostic medical tests are used to forecast outcomes of interest. A performance study includes result X from the test and reference result Y for the outcome. Sometimes X or Y is measured with error by surrogate W, complicating performance evaluation of the test. Many examples exist in which reference result Y is measured with error, leading to so-called verification bias. Other examples exist in which X is measured with error. E.g., a test yielding result W may be FDA approved based on a study that includes Y, and X may be the result from an unstudied new test with the same intended use. Or, W may be the result from a clinical trial assay used to select patients for a therapeutic trial, while X is the result from a market ready test, available only after the trial is over. Performance evaluation in the main study can be adjusted for measurement error using supplemental information, including verification of Y on a subset of subjects, replicate measurements of W, or external calibration of W to X. For the therapeutic trial, we'll use Bayesian methods to impute missing test result X from surrogate W, outcome Y, and covariate values Z based on external calibration data.