11.10.2020 Model Building Training
Presenter: Kemal Oflus
Watch Recording | PPT
What is a statistical model?
Statistical models are used to quantify the relationship between one or more factors (inputs, X's) and a response (output, Y). Your modeling goal might be to explain or predict. A statistical model can help you:
- identify the most important factors (explain)
- understand how the response changes as a function of the factors (explain)
- quantify the variability in the response that is explained the factors and their interactions. (explain)
- make predictions about future outcomes (predict)
What is model overfitting?
An overfit model is a statistical model that is more complex than it needs to be. Overfitting is modeling the noise in the data, rather than the true relationship between the factors. Consequence of overfitting: Your model isn't going to predict well and may place too much trust in an unreliable prediction or result...leading to increased cost, waste, missed deadlines, etc.
When should you use statistical modeling?
- You spend a lot of time dealing with unanticipated problems.
- Your problem-solving efforts yield unpredictable returns – and the same problem or a related one reoccurs.
- You feel you don’t have the complete picture and have to make decisions based on incomplete information.
- You transfer products and processes with limited understanding of how they work.
- You feel stuck in a vicious circle where managing existing problems or processes limits your time for optimization, innovation and new development.
What are the benefits of statistical modeling?
- Reduce number of learning cycles.
- Reduce total problem-solving time.
- Increase predictability of improvement and problem-solving activities.
- Increase useful understanding (creating incomplete knowledge/decisions from your data may result in false solutions)
- Make sustainable decisions within the required time frame.
- Transfer more complete process knowledge to upstream colleagues.
- Increase time available for new projects to grow future opportunities because it is less likely the same problem comes back to be solved again (and again).
What are the most common types of models?
|
| Type |
What It Does |
Helps Users to |
| Multiple Linear Regression |
- Predicts the average response value based on values of multiple predictors/factors.
- Modeling type is continuous (ex: dollars spent)
|
- Identify and optimizing critical to quality characteristics, with the goal of developing low cost and high quality products
- Predict customer spending based on demographics information and historical buying patterns
- Develop pricing strategies based on product mix and consumer characteristics
- Predict future behavior from the results of experiments
- Predict yield
|
| Logistic Regression |
- Model the probability that an event will occur.
- Modeling type is categorical
- nominal - pass/fail
- multinomial - red, blue green
- ordinal - small, medium, large
- Commonly used because you don't have continuous response data
|
- Identify potential fraud
- Determine the causes of defective items in an assembly line
- Understand reasons behind flight risk for employees
- Determine the likelihood of purchase for a customer on a website
- Pass/Fail quality inspection
|
Decision Trees
(webcast)
|
|
|
Neural Networks
(webcast)
|
|
|
| How do different industries use models? |
| Industry |
Typical Problems trying to solve with Models |
Semiconductor
(Semiconductor resources page) |
- Yield Management
- Defect Modeling
- Defect Clustering
|
Pharmaceutical
(Pharmaceutical resources page) |
- Research and Development
- DOE Optimization
- Biomarker Analysis
|
Chemical
(chemical resources page) |
- Mixture Optimization
- Preventative Maintenance
- Yield Management
|
Consumer and Product Goods
(CP& G resources page)
|
- Finding the best set of customers
- Customer Lifetime Value (CLV) calculations
- Sales Forecasting
- Market Research
|
Model Building in JMP
|
Value Statement: JMP users are solving more scientific, engineering and business problems correctly the first time – and in less time – by extracting powerful insights from existing data using proven, simple statistical modeling methods - without needing a strong statistical background and without having to use multiple platforms or products.
|
| Modeling Options in JMP |
Discovery Questions for Modeling in JMP |
- Multivariate Linear Regression
- Basic Decision Trees
- Simple Neural Networks
Used in conjunction with modeling:
- Principle Components Analysis
- Clustering
|
What problems are you trying to solve with your data?
Do you face challenges that can’t be solved by an appeal to expert judgment, simple spreadsheet analysis, or by varying only one thing at a time?
Are you forced to regularly make decisions from incomplete data or information, limiting your understanding of the real drivers of robust, effective and efficient processes?
What is your pain points and how much is that costing (both in time and money)? What is the cost of getting it wrong?
How many tools are you currently using to analyze your data?
How often do you analyze data?
- Are you having to rerun analyses on a regular basis?
- Does the need to fix problems with existing processes limit the time you can spend on innovating and developing new processes and products
- How important is it for you to have a good understanding of the process?
|
| Model Building in JMP Pro |
Value Statement: JMP Pro enables our users to explore and understand data with an easy-to-use interface by providing a collection of the most relevant statistical instruments for analyzing data without being overwhelmed by choices and without needing to program. What JMP Pro delivers is TIME and CONFIDENCE. The addition of Model Screening (JMP 16) and Formula Depot allows modelers to stay inside a single toolkit, and do something VERY quickly, that typically takes a great deal of effort. Validation delivers confidence.
|
| Modeling Options in JMP Pro |
Qualifying Questions to see if they are good Pro Candidate
|
|
- Ask any technical question. Example: What kind of R-squares are you getting now out of your models?
- If they say .99, you know JMP Pro might not be as impactful compared to someone who says 0.30.
- Do you use cross validation or stepwise regression?
- In Pro you can use Cross-validation to prevent over-fitting a model which is a huge advantage for JMP Pro. If they have an understanding of cross validation or stepwise regression that is typically a good sign they are a Pro candidate.
|
More Modeling Resources to Share:
