In the final session of JMP On Air’s Predictive Modeling Day, Director of JMP Research and Development Chris Gotwalt gives us his machine learning tips and shows us how he builds successful models in JMP. After a brief overview that compares classical modeling against machine learning modeling, Chris discusses how he sets up data for successful modeling, how he uses visualizations to identify good models and diagnose problems, and how to use the ‘Partition Trick’ to solve rare event classification problems.   ... View more

Die Analyse von funktionalen Daten und deren Verknüpfung mit den untersuchten Faktoren wurde mit dem Functional Data Explorer (FDE) in JMP wesentlich vereinfacht. Wenn die Ergebnisse von Versuchen einfache Messwerte sind, kann man in direkter Weise ein Modell anpassen und damit Einstellungen der Faktoren finden, die die Zielgröße die gewünschten Spezifikationen erreichen lässt. Aber wie kann man erreichen, dass eine Idealkurve getroffen wird? Der FDE bietet nun (in JMP 15) eine Lösung, die mit einem Klick auszuführen ist. Eine ideale Kurve anzunähern ist jetzt vielleicht noch einfacher als es bisher war, einen Zielwert zu treffen.   Statistische Prozesskontrolle ist ein integraler Teil des Qualitätsmanagements und es gibt gute Gründe, dabei auch multivariate Regelkarten einzusetzen. Das wird mit JMP 15 viel einfacher und leistungsfähiger. Diese Plattform verbindet Dimensionsreduktion und drill-down Funktionalität mit multivariaten Regelkarten. Sie zeigt den Verlauf der T² Statistik ebenso wie den der quadrierten Fehlersummen. Für jeden Messpunkt kann man im Detail die Beiträge der einzelnen Prozesse sehen, die jeweiligen Kontrollkarten bis hinunter zur einzelnen Beobachtung. Hier eröffnet sich eine neue Qualität der Regelkarten. ... View more

Level: Advanced   Chris Gotwalt, JMP Director of Statistical Research and Development, SAS Ryan Parker, JMP Senior Research Statistician, SAS   JMP Pro is now able to model and visualize functional data in a new way that is direct, straightforward and leads to highly accurate models. We will introduce the Functional Data Explorer platform and go through a set of examples that illustrate two particularly useful families of problems that it facilitates solving, function response DOE analysis (FR-DOE) and functional regressor machine learning (FR-ML). FR-DOE analysis allows one to see directly how the shape of a response curve changes as a result in changes in the value of the DOE factors. FR-ML is an approach to feature extraction for sensor data from manufacturing processes that faciliates accurate early prediction of final batch yield or faulty products. Our examples will be primarily from pharmaceutical, chemical and semiconductor manufacturing, but our approach can be applied to practically any industry. ... View more

Level: Advanced Job Function: Analyst / Scientist / Engineer Chris Gotwalt, JMP Director of Statistical Research and Development, SAS In this tutorial we will give a general overview of how to apply machine learning methods using the tools in JMP Pro. The essential idea is that the basic tools of machine learning actually make modeling easier than earlier approaches to modeling data. We will begin by introducing the basic concepts of how models are chosen in a machine learning framework, then show how to set that up with the Make Validation Column utility. Then we will demonstrate ADI, a new approach to missing data imputation that has been added to JMP Pro 14. After that we will illustrate the most important platforms for machine learning in JMP Pro: Neural, Partition and Generalized Regression. At the end we will use Model Comparison to make a final decision about what model to use and use the Formula Depot to create scoring code to use the model outside of JMP Pro. Throughout we will illustrate the unique interactivity in JMP and its close association between data and graphs, which makes the modeling decision process intuitive and transparent, while also also letting one "see inside the black box" of machine learning models in a way that no other product can. ... View more

Level: Advanced Job Function: Analyst / Scientist / Engineer Chris Gotwalt, JMP Director of Statistical Research and Development, SAS Ryan Parker, JMP Senior Research Statistician, SAS JMP Pro is now able to model and visualize functional data in a new way that is direct, straightforward and leads to highly accurate models. We will introduce the Functional Data Explorer platform and go through a set of examples that illustrate two particularly useful families of problems that it facilitates solving, function response DOE analysis (FR-DOE) and functional regressor machine learning (FR-ML). FR-DOE analysis allows one to see directly how the shape of a response curve changes as a result in changes in the value of the DOE factors. FR-ML is an approach to feature extraction for sensor data from manufacturing processes that faciliates accurate early prediction of final batch yield or faulty products. Our examples will be primarily from pharmaceutical, chemical and semiconductor manufacturing, but our approach can be applied to practically any industry. ... View more

Level: Intermediate Chris Gotwalt, JMP Director of Statistical Research and Development, SAS Philip Ramsey, Principal Lecturer, University of New Hampshire; and Owner, North Haven Group There are two different goals to statistical modeling: explanation and prediction. Explanatory models often predict poorly (Shmueli, 2010). Often analyses of designed experiments (DOE) are explanatory, yet the experimental goals are prediction. DOE is a best practice for product and process development where one predicts future performance. Predictive modeling requires partitioning the data into training and validation sets where the validation set is used to assess predictive models. Most DOEs have insufficient observations to form a validation set precluding direct assessment of prediction performance. We demonstrate a “balanced auto-validation” technique using the original data to create two copies of that data, one a training set and the other a validation set. The sets differ in row weightings. The weights are Dirichlet distributed and “balanced”; observations contributing more to the training data contribute less to the validation set (and vice versa). The approach only requires copying the data and creating a formula column for weights. The technique is general, allowing one to apply predictive modeling techniques to smaller data sets common to laboratory and manufacturing studies. Two biopharma process development case studies are used for demonstration. Both cases have large validation sets combined with definitive screening designs. JMP is used to demonstrate analyses.   ... View more

JMP 11 is a big step forward on many fronts, but especially for JMP Pro. From the time it was introduced in JMP 9, JMP Pro has been predominantly focused on predictive modeling, with more sophisticated procedures for fitting neural networks and boosting and bootstrapping regression tree models. JMP Pro 11 broadens its scope substantially with the addition of four new platforms: Fit Mixed, Generalized Regression, Uplift, and Reliability Block Diagram. The Fit Mixed platform was developed by Laura Lancaster, John Sall and myself and brings to JMP Pro the ability to fit linear mixed models with different correlation structures that take into account space and time, random coefficients models among others. The platform is implemented using sparse matrices and is multithreaded, allowing you to fit models to larger data sets with fairly large numbers of random effects. You can take advantage of these models in almost every scientific discipline, from the social and environmental sciences to engineering and manufacturing. The Generalized Regression platform developed by Clay Barker is our next-generation generalized linear modeling platform. As the successor to the Generalized Linear Models platform, it can fit models to non-normally distributed data. The most powerful aspect of the platform is the incorporation of modern regression tools like the Lasso and Elastic Net, which automate the process of model fitting and variable selection. The elastic net, in particular, is very useful when fitting models to observational data with correlated inputs. In the future, this will be where we will be adding more distributional modeling options, as well as things like Generalized Estimating Equations and possibly even Bayesian Hierarchical models. The Reliability Block Diagram platform developed by our Principal Reliability Developer, Peng Liu, in collaboration with the author of JMP’s NodeGraph library, Aaron Anderson, rounds out our suite of engineering reliability tools. It enables one to predict the reliability of complicated products made of many components. For example, an automotive manufacturer will have performed preliminary experiments to understand reliability of the various subcomponents of a new car that hasn’t come out yet. Using the Reliability Block Diagram platform, you can combine these reliability models for subcomponents – such as the transmission, powertrain, engine and electrical system – with an easy-to-use graph-based interface that then plots the reliability distribution for the overall car. The Uplift Modeling platform, developed by John Sall himself, is a powerful data mining tool with applications that vary from direct marketing to the pharmaceutical industry. The platform fits models that let you identify subpopulations that are most likely to respond to a treatment. For example, you could identify which people are most likely to make a purchase after including them in a direct marketing campaign or identify the subpopulations of patients that are most likely (or unlikely!) to respond to a drug treatment. We’ve also kept very busy working on the standard edition of JMP. Possibly the most significant effort on this front has been our renewed commitment to serving folks in the social sciences. The centerpiece of these efforts has been another project led by John Sall: a dramatic overhaul of the Categorical platform to make it easier to analyze market research surveys. Factor analysis, an essential part of the social scientist’s statistical toolkit, has been given its own JMP platform by our multivariate analysis developer, Jianfeng Ding. JMP 10 initiated a renewal of the quality and statistical process control components in JMP and saw the introduction of the Control Chart Builder and the Measurement Systems Analysis (MSA) platform. Developed by Annie Zangi, the Control Chart Builder is a drag-and-drop tool for creating Shewhart charts for continuous response data that has been expanded to handle attribute charts for count and categorical data, as well as charts for monitoring the behavior of rare events. Laura Lancaster’s MSA platform is the first commercial implementation of Don Wheeler’s EMP approach that is an alternative to Gauge R&R studies. The nice thing about this is the simple and direct way it makes clear the impact of measurement error on the ability to detect shifts in a manufacturing process. In JMP 11 the MSA platform has been expanded to handle more sophisticated models that more realistically describe how measurements are taken from a modern manufacturing process. One of the features that I am personally most excited about is Ryan Parker’s addition to the Profiler in JMP. Ryan has developed a simulation-based approach to assess variable importance that is based on only a prediction formula and not on how you fit the model. The technique is based on the Sobol decomposition of a function, and you can apply it to just about any formula. What this means is that you now have a tool that can be used for variable selection that can be applied to any statistical model, even ones like neural networks that are too much like a black box to use traditional variable selection methods. Here’s the Variable Importance tool applied to a two-layer neural network model fit using JMP Pro to some (simulated) credit card approval data. Another cool new widget in JMP is the Multiple Comparisons module developed by Laura Lancaster that makes it easy to do a wide variety of multiple comparisons procedures on arbitrary sets of combinations of input variables. These include Analysis of Means, Tukey’s all pairwise comparisons, Dunnett’s test for comparisons versus a control, and even equivalence testing when there is a threshold for practical significance. Each of these comes with a nice graph that makes it easy to see where the differences are in your data. Below is an example of the one for equivalence tests: JMP 11 really has a lot for you to look forward to! Note: This post is part of series of previews of JMP 11 by JMP R&D members and technical experts. The JMP Blog will feature a new preview every Tuesday until JMP 11 is released in September. Check out all of the JMP 11 previews. ... View more

It’s almost time for JSM again, which is always one of my favorite times of year! The JMP team is having another big year at JSM with 10 presentations, panel discussions and poster sessions, not to mention the second annual JSM JMP Mixer. Last year’s JMP Mixer in San Diego was a great success and was certainly a highlight of the conference for me. I enjoyed catching up with my friends in the JMP community. This year, the JMP Mixer will be held Sunday, Aug. 4., from 6:30-8:30 p.m. in the Fortifications Ballroom of Le Westin Montréal. Fortunately for me, I’m "off the hook" as far as having to give a presentation this year, but there are two presentations I have collaborated on, at least to some degree. The first one is on Monday, Aug. 5, and is being given by Elizabeth Claassen, one of our summer research interns.  She’ll be presenting preliminary simulation results related to a bias adjustment we have developed for variance components in binary response random effects models. Walter Stroup, her Ph.D adviser at the University of Nebraska, Elizabeth and I have derived a Firth biased adjusted estimation procedure that is intended to have the same kind of reduced bias properties in the generalized linear mixed model (GLMM) that REML provides for the linear mixed model. On Wednesday, Aug. 7, JMP colleague Laura Lancaster will be presenting the results of an investigation of role of sample size and study design for Gauge R&R-type measurement systems analysis experiments. Specifically, she will be discussing how Gauge R&R study design affects the precision with which one can predict a measurement system’s ability to correctly identify shifts in a manufacturing process. Hope to see you in Montreal! Here is a complete list of all the presentations given by JMP folks here at SAS: Sunday, Aug. 4, 4:05 p.m., Room CC-524b, Presentation Recent Development in Biomarker and Subgroup Identification for Tailored Therapeutics, Russ Wolfinger and Richard Zink Monday, Aug. 5, 2:00 p.m., Room CC-220bc, Poster Session Dynamic Exploration of the Soul of the Community, Sam Gardner, Nicole Jones, and Michael Crotty Monday, Aug. 5, 2:35 p.m., Room CC-513a, Presentation Bias Correction for Covariance Parameter MLEs in GLMMs, Elizabeth Claassen and Chris Gotwalt Tuesday, Aug. 6, 11:50 a.m., Room CC-518, Presentation Optimal Split-Plot Designs for Fixed-Effect and Variance-Component Estimation, Brad Jones (Peter Goos presenting) Tuesday, Aug. 6, 2:00 p.m., Room CC-524a, Panel Discussion A Celebration of J. Stuart Hunter's Contributions to Statistics, Brad Jones (panelist) Tuesday, Aug. 6, 3:35 p.m., Room CC-512h, Presentation SPC Data Visualization of Seasonal Data, Annie Zangi and Di Michelson Wednesday, Aug. 7, 3:05 p.m., Room CC-520a, Presentation Exploring Measurement System Study Sample Size and the Power to Detect Production Process Shifts, Laura Lancaster and Chris Gotwalt Wednesday, Aug. 7, 3:20 p.m., Room CC-520a, Presentation Interactive SPC: A Textile Quality Case Study, Scott Wise Wednesday, Aug. 7, 3:20 p.m., Room CC-518, Presentation A New Derivation and New Perspective of Weibayes,  Peng Liu Wednesday, Aug. 7, 3:00 p.m., Room W-Ville-Marie, Continuing Education Course Using the Bootstrap Feature in JMP, Michael Crotty and Clay Barker ... View more

It’s that time of year again -- we’re all gearing up for the Joint Statistical Meetings (JSM), this time in San Diego. JSM week has always been one of my favorite times of year. For us JMP R&D folks, JSM is a great change of pace: We put our sales hat on for a few days and meet customers at the JMP booth. It’s also an excellent opportunity for us to see what’s going on in the statistical research community and for us to present the things we have been working on over the last year. I always enjoy being able to spend time with my JMP colleagues who work around the country -- and with some just down the hall who I never get to see other than in meetings and the occasional lunch. Even more than usual, this year I’m looking forward to JSM. For the very first time, JMP has its own JSM reception. The reception is open to all JSM attendees and is on Sunday, July 29, from 7 to 9 pm in Aqua Room 300 at the Hilton Bayfront. We hope to see you there. We also have a great lineup of presentations, chaired sessions, short courses and panel discussions this year, with a record 10 folks from JMP contributing in one way or another: Russ Wolfinger, Richard Zink and Walter Boyle, "Dynamic Comparison of Simulated Adaptive Trials," 4:05 pm, Sunday, July 29, CC-Room 28D Bradley Jones, "Efficient Designs with Minimal Aliasing," 9:15 am, Monday, July 30, CC-Room 31C Laura Lancaster and Christopher Gotwalt, "Exploring the Relationship Between Your Ability to Detect Production Process Shifts and Measurement System and Process Settings," 2:05 pm, Monday, July 30, CC-Room 29A Don McCormack, Session Chair, "Methods of Reliability and Life Testing," 8:30-10:20 am, Tuesday, July 31, CC-Room 30A Peng Liu and Michael Crotty, "Determine Drug's Expiration Date: Some Interesting Findings," 8:35 am, Tuesday, July 31, CC-Room 30A Christopher Gotwalt, "What is REML?" 8:50 am, Tuesday, July 31, CC-Room 31C Russ Wolfinger and Richard Zink, Short Course CE_28T, "Predictive Modeling in the Life Sciences," 8:00-9:45 am, Wednesday, Aug. 1, HQ-Indigo 202T T. Beasley (UAB), Russ Wolfinger, Wendy Czika, Hemant Tiwari (UAB), and David Allison (UAB), Poster Presentation, "Using the Polynomial Power Method to Simulate Data with Correlated Non-Normal Error Structure for Assessing Methods Proposed for Family-Containing Genetic Association Tests," 10:30 am-12:20 pm, Wednesday, Aug. 1, CC-Sails Pavilion Curt Hinrichs, Business Analytics Panel Discussion, 8:30 am, Thursday, Aug. 2, CC-Room 31B Clay Barker, Leonard Stefanski (NCSU), Jason Osborne (NCSU), "A Resampling Approach to Degree-of-Freedom Estimation," 11:50 am, Thursday, Aug. 2, CC-Room 24B ... View more