Distilling greater value from data

Accessible statistical support and consultancy

Call: + 44 (0) 1732 465662

INTRODUCING

Correlated Component Regression

A better way to build predictive models

Correlated Component Regression (CCR) is a revolutionary new approach to Predictive Modelling which is set to take the statistical world by storm. Developed over several years by Dr Jay Magidson, founder of Statistical Innovations and pioneer of Latent Class Analysis, this is more than just a software application; it is a completely new way of model building consistently outclassing conventional methods.

Gary Bennett became involved as beta tester for the project in late 2010 after attending one of Dr Magidson's seminars. As a statistical consultant of some 20 years in the Marketing Research Business he immediately saw the potential benefits of this new approach. Many of Gary's end-clients have since benefited from its use including leading media organisations, retailers, political parties, IT, Telecoms, Pharmaceutical and Financial Services companies.

CCR was originally recommended as a modelling approach for high dimensional data sets with few cases but lots of predictors (or in extreme cases more predictors than cases). However as well as out-performing other methods on high dimensional data it is also best in class for regular data sets.

Like many of the best innovations in statistical modelling it has a simple and transparent set of principles at its core:

	Conventional Predictive modelling	Correlated Component Regression (CCR)
Sample requirements	Large samples / large sample to predictor ratio - model assumptions break down if this is violated	Works with what is available (used on data sets with as few as 20 cases)
Model Selection	Based on hypothesis testing and large sample assumptions i.e. how well model SHOULD perform on new cases under IDEAL sample conditions	Based purely on cross-validated performance i.e. how well model REALLY performs on new cases
Predictor Correlations	Multicollinearity makes hypothesis testing less robust and can produce counter-intuitive coefficients	Can appropriately handle even severe multicollinearity & uses a stabilising algorithm which produces more robust coefficients
Over-fitting Noisy data	Tendency to over-fit noisy data	Designed specifically to MINIMISE overfitting and MAXIMISE ability to predict new cases
Model Simplicity	Tendency (especially with large samples) to result in OVERLY COMPLEX model	Tends to yield SIMPLE models with few terms which generalise well to new cases
Interpretability	Models often difficult to interpret	Predictors load onto virtual correlated components (factors) which can be interpreted to understand the dimensions of the prediction
Variable selection	Stepwise methods which rely on statistical testing and tend to "capitalise on chance�	Stepping Down procedure based on size of standardised (and stabilised) coefficients
Suppressor Variables	Often the most important predictors and missed by many standard procedures	Always ensures these are captured in the model

In summary CCR:

• Maximises Out-of-Sample Performance (ability to predict new cases)
• Produces simpler, more interpretable models
• Enables very robust models to be built with minimum demands on the data
• Out-performs other state-of-the art methods

The intuitive appeal and simplicity of this approach, combined with the hard facts makes this a first choice method for analysts working with REAL data. If you care about your models this is a must have method for your statistical tool-kit. We strongly believe that in years to come this will be seen as a GAME CHANGING methodology.

Our CCR Methodology Overview gives a more in-depth description of the algorithm & comparisons with other approaches to modelling "high dimensional data". Our Library of Technical Papers contains further publically available technical information on the methodology

	CCR: A Brief Methodological Description (Magidson and Bennett, 2011) This note contains a brief technical overview of the CCR Methodology
	Library of Technical Papers Publically available Technical Papers on CCR

---

The original implementation of Correlated Component Regression from Statistical Innovations

Download FREE Demo version
CORExpress Pricing
License Purchase
License Renewal

Why buy CORExpress^® from Logit?

Logit Research is Statistical Innovations exclusive UK-based partner for CCR and reseller for CORExpress.

Logit Research's core activity is statistical consultancy. We are in business to provide access to state-of-the-art modelling, but with a large dose of common sense, pragmatism and plain speak.

We do not just exist to sell software; we want to promote this family of methods and help other statisticians, analysts and researchers break through the dogma of conventional approaches to model development.

Our extensive experiences of working with CORExpress on real survey research projects put us in an ideal position to provide both routine and extended support to clients as and when needed.

Together with the commitment, professionalism and provision of sound advice for we are known for in the Marketing Research business this makes us an excellent choice of partner and distributer.

Being in the same Time zone also makes us a very convenient partner for users in the UK and Europe.

Purchasing a license direct from us will provide you with:

• On-going maintenance and support for CORExpress. Email support for Routine questions about the CORExpress implementation of CCR is included.
• All upgrades to CORExpress covering the period of the license.

Extended support is also available on an ad-hoc basis for specific applications. Please address enquiries to support@logitresearch.com.

 

---