Comparison of methods for robust parameter design of products and processes with an ordered categorical response

Bacanlı Karabulut, Gonca
Robust design of products or processes with categorical response has more momentous role in industrial experiments for quality improvements, because ordinal categorical quality characteristics are encountered more frequent than continuous ones in industry. In this study, five optimization methods for an ordered categorical response are compared with each other: Logistic Regression Model Optimization (LRMO), Accumulation Analysis (AA), Weighted Signal-to-noise Ratio (WSNR), Scoring Scheme (SS), and Weighted Probability Scoring Scheme (WPSS). In order to compare performance of these methods for different types of robust design problems, each method is individually applied on two different types of problems: smaller-the-better and larger-the-better. All examples are studied to find the optimal parameter settings of statistically significant controllable factors and trying to optimize both location and dispersion of the results. To compare the optimal levels derived from these five methods, three performance criteria are used: SNR at optimal parameter settings, estimated by ANOVA model of continuous version of data or true model (if applicable); probability of observing target category, estimated by LR models; and that of observing target category estimated by ANOVA models of cumulative percentage of categories. According to the results, LRMO and AA methods have the best performance results in most of the examples analyzed in this study. Since AA is criticized as not allowing analysis of location and dispersion effects separately but not LRMO, WPSS and SS, more examples and further analysis might be studied to show this discrepancy of the methods.