Online Program

A Comparison of Standard Logistic Regression, Multilevel Modeling, Robust Error Estimation, and Exposure Simulation for Data Containing Quasi-Berkson Error (303578)

 View Presentation

Keywords: quasi-Berkson error, measurement error, partially ecologic data

In studies of workplace injury, such as carpal tunnel syndrome (CTS), the standard ways of measuring risk factors are expensive to collect or biased. Yet, there are publically available data sets containing summarized ratings of different exposures by job title. When combined with data containing worker injury and other individual level factors, the merged data have a quasi-Berkson error structure. Using standard logistic regression to analyze data with this error structure produces biased results. Thus, others have recommended using multi-level modelling or robust error estimation to adjust for biased variance when analyzing data with this structure. We sought to compare the performance of standard logistic regression, multilevel modeling, robust error estimation, and exposure simulation in realistically simulated data. Several sets of simulated data were created with varied assumptions. Models were run on these data. Bias, coverage, convergence, and power were compared between approaches. All approaches produced biased results. The standard logistic regression approach had the best overall performance. The robust error approach had similar results.