Return to JMP Fed Gov User Resources Page
22 Dec Webcast - Strategies for Analyzing DOEs
In this webcast we will analyze several DOE data sets. For visual analysis tools we will use the Distribution and the Graph Builder, including data filtering on columns of factors, as well as stacking of factors. Modeling strategies will include conservative approaches such as looking at first order effects before moving on to second order effects guided by "effect heredity" and "effect sparsity" principles. Aggressive strategies will include stepwise regression using several different stopping criteria to prevent overfitting and even fitting "All Possible Models." Actual vs. Prediction plots with checkpoints can be used to help choose models. Various sample graphical and stepwise regression output that may be recreated in the session are shown below:
Distribution Platform showing shading of factor levels with top half of response data selected.
![7788_Distribution.jpg 7788_Distribution.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1072i01F118E65929F1B0/image-size/medium?v=v2&px=400)
Graph Builder Yield vs Factor Ranges for all 10 Factors. Data summarized by Mean with error bars equal to the confidence interval about the Mean. Smoother curve (blue) and Line Fit (red) to data. Relative size of Main Effects and Curvature can be seen.
![7805_GB 2X5 for 5 factors.jpg 7805_GB 2X5 for 5 factors.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1073i50D8142B762FDE33/image-size/medium?v=v2&px=400)
Graph Builder showing response values plotted vs individual factor settings with "Smoother" curve
![7789_GB Yield vs A and B.jpg 7789_GB Yield vs A and B.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1074iE2528081AD1C6F61/image-size/medium?v=v2&px=400)
Graph Builder showing response vs factor settings of one factor with second factor overlayed on graph.
Non-parallel lines are indicative of an interaction between factors.
![7790_GB Yield vs B overlay A.jpg 7790_GB Yield vs B overlay A.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1075iE94B72D86D215371/image-size/medium?v=v2&px=400)
Plot of criterion history for stepwise regression of 24 observations choosing a 4-factor, 9-term model subset of a 10-factor, 66-term quadratic model that is "over parameterized."
![7791_Criterion History.jpg 7791_Criterion History.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1076i25A6027634BD8DAE/image-size/medium?v=v2&px=400)
Chart of sorted parameter estimates showing dominance of factors A & B
![7792_Scaled estimates.jpg 7792_Scaled estimates.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1077i2C7AC8B9A9CB0EF7/image-size/medium?v=v2&px=400)
Prediction Profiler set to conditions predicting maximum yield withing ranges of factors. Three checkpoints yielded values of 15.10, 15.93 and 16.16 - all within 95% confidence window (14.4, 20.1).
![7902_CO2 Profiler.jpg 7902_CO2 Profiler.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1078i61675DB50135595D/image-size/medium?v=v2&px=400)
Best 1-term through 8-term models (ignoring constant) from fitting all 10 million possible models for 8 factors.
NOTE domination by A & B followed by C and then suggestion that G may be important.
![7793_All Possible Models up to 8 terms.jpg 7793_All Possible Models up to 8 terms.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1079iF943DBD5B657BD8D/image-size/medium?v=v2&px=400)
Overlay Plot of 4 metrics from an All Possible Model Table for factors A, B, C, E, and F;
RSquare, RMSE, AICc and BIC vs Model Terms.
![7806_Overlay Plot All Model Table Metrics.jpg 7806_Overlay Plot All Model Table Metrics.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1080i175B70B9607BF142/image-size/medium?v=v2&px=400)
Plot of Actual vs Predicted for 3-factor model fit to 24 design trials. Four Checkpoints NOT used in fit are also plotted.
![7794_Act vs Pred with Checkpoints.jpg 7794_Act vs Pred with Checkpoints.jpg](https://community.jmp.com/t5/image/serverpage/image-id/1081i8A5CC3DD3965D5A1/image-size/medium?v=v2&px=400)