Statistical Significance Testing as a Guide to Partial Least-Squares (PLS) Modeling of Nonideal Data Sets for Fuel Property Predictions

Partial least-squares (PLS) was used to formulate property models for a set of 43 jet fuel samples using near-infrared (NIR), Raman, and gas chromatography (GC) data. A total of 28 different properties were evaluated for each technique. Given that the data set was small and several of the property distributions were nonideal, significance testing was used for model formulation and evaluation. A statistical F test was applied for selection of latent variables, determining the significance level of the model compared to random chance and evaluating the increase in prediction error when the expected sources of error were fully incorporated. A chart was used to categorize the confidence in the modeling ability as high, low, or indeterminate with the given data set.