Prev
Correlated Component Regression: A New Method for Analyzing Big Data or Small Data Containing Many Correlated Predictors
Presenter: Jay Magidson, PhD (Statistical Innovations Inc.)
Thursday, September 27, 11:00 am - 1:00 pm Eastern time
Description
In this webinar, we introduce a new regression method - called Correlated Component Regression (CCR), which provides reliable predictions of a dependent variable even with collinear or near multicollinear data. Different variants of CCR are tailored to different types of regression (e.g., linear, logistic, Cox regression). CCR also includes a step-down variable selection algorithm for eliminating irrelevant predictors. Unlike PLS and penalized regression approaches, CCR has the favorable property of scale invariance, resulting in the same predictions whether based on unstandardized or standardized predictors (z-scores).
Following a discussion of multicollinearity problems and use of M-fold cross-validation to avoid overfitting in model development, CCR is introduced and applied using the CORExpress® software in several examples involving real and simulated data. These data are made available to all attendees.
The examples are:
1. Linear regression with P=6 correlated predictors and a small sample size (N=24)
2. Linear regression with P=700 Near Infra-Red (NIR) wavelengths to predict fat content
3. Simulated data based on gene expression data according to the assumptions of Linear Discriminant Analysis (LDA)
4. Predicting Liking of several orange juice products based on many correlated product attributes
A detailed 2-day course is also being offered in Boston as part of Statistical Modeling Week. For full details see: statisticalinnovations.com/workshops/workshops.html .
Presenter Bio
Jay Magidson is founder and president of Statistical Innovations Inc., a Boston based consulting, training and software development firm specializing in innovative applications of statistical modeling. He taught statistics at Tufts and Boston University and is widely published on the theory and applications of multivariate statistical methods. His consulting clients have included Pfizer, Chevron, AC Nielsen, and Kellogg's. Dr. Magidson designed CORExpress®, XLSTAT-CCR, SPSS CHAID, GOLDMineR®, and is the co-developer (with Jeroen Vermunt) of the Latent GOLD® and Latent GOLD® Choice programs.
Registration Deadline: Tuesday, September 25
Registration Fees
- Members of the Q&P and SPES Sections: $75
- ASA Members: $90
- Nonmembers: $105