Abstract:
|
To characterize the capability of an aircraft inspection system, hit and miss indications from the system must be collected over a range of crack sizes. Insufficient overlap or uneven counts of hits and misses may cause separated or quasi-separated classes, which cause problems for logistic regression modeling because the maximum likelihood estimator may fail to converge to a unique solution. Resulting parameter estimates are unavailable or unreliable. Extensive simulations of representative Lognormal, Weibull and Uniformly distributed data were used to test the quality of the logistic regression model and three alternative models: Firth's Bias Adjusted Likelihood, the Lasso, and a ranked set resampling method from nonparametric statistics. The quality of each type of model was assessed by comparing the existence of an estimate and the relative percent bias within any existing estimates, with respect to overlap, evenness, and sample size.
|