Abstract:

Probability of Detection (POD) studies using the hit/miss criteria often suffer from quasiseparation in the data, which leads to a lack of convergence for the maximum likelihood equation, making all results from the logistic regression questionable. To find a converging solution, three techniques showed the most promise: the Lasso, Firth's biasadjusted likelihood technique, and bootstrapped ranked set sampling, a nonparametric resampling technique known to minimize separation. These methods were applied to data which suffers from quasiseparation, specifically, a probability of detection study for finding flaws in aircraft fastener heads using eddy current. The resulting parameter estimates and POD estimates will be compared and the usefulness of each method to overcome separation issues demonstrated. The overall goal and motivation for the use of this technique is to be able to accurately predict in postprocessing the smallest flaw size an inspector could miss with the sensor even in the case of quasiseparated data.
