Getting started with Correlated Component Regression (CCR) in XLSTAT-CCR

Dataset for running Correlated Component Regression

This tutorial is based on data provided by Michel Tenenhaus and used in Magidson (2011), “Correlated Component Regression: A Sparse Alternative to PLS Regression”, 5th ESSEC-SUPELEC Statistical Workshop on PLS (Partial Least Squares) Developments.

The data consists of N=24 car models, the dependent variable PRICE = price of a car, and P=6 explanatory variables (predictors), each of which has a positive correlation with PRICE

but each predictor also has a moderate correlation with the other predictor variables

CCR will apply the proper amount of regularization to reduce confounding effects of high predictor correlation, thus allowing us to obtain more interpretable regression coefficients, better predictions, and include more significant predictors in a model than traditional OLS regression.

The OLS regression solution maximizes R² in the training sample, yielding R²= .85. However, since this solution is based on a relatively small sample (N=24) and correlated predictors, it is likely that this model overfits the data and that .85 is an overly optimistic estimate of the true population R². Consistent with an overfit model, Table 1 shows that the OLS solution yields large standard errors and unrealistic negative coefficient estimates for the predictors CYLINDER, SPEED, and WIDTH.

Table 1: Results from traditional OLS regression: CV-R² = 0.63

Moreover, POWER is the only predictor that achieves statistical significance (p=.05) according to the traditional t-test.

CCR utilizes the cross-validated R² as its criterion for determining the proper amount of regularization (K) to use in a regression model. Fig.1 shows that substantial decay in CV-R² occurs for K>2. Thus, a substantial amount of regularization is required (K2 = .75 for CCR compared to .63 for OLS regression.

Fig. 1. Cross-Validation Component (CV-R²) Plot showing deterioration for K>2

Also, in contrast to OLS regression which yields some negative coefficient estimates, CCR yields more reasonable positive coefficients for all 6 predictors as shown below.

Table 2. CCR unstandardized / standardized coefficients with K=2 components.

The first part of this tutorial shows how to use XLSTAT-CCR to obtain these results.  The second part (see ‘Activating the Step-down Algorithm’) shows how to activate the CCR step-down procedure to eliminate extraneous predictors and obtain even better results (CV-R² = .77) as indicated in the following table.

  Table 3. Results from CCR with step-down algorithm

To activate the Correlated Component Regression dialog box, first start XLSTAT by clicking on the  button in the Excel toolbar, then select the XLSTAT / Modeling data / Correlated Component Regression command in the Excel menu or click the corresponding button on the Modeling data toolbar.

Once you have clicked the button, the Correlated Component Regression dialog box is displayed with the Method=CCR.LM (linear regression model) selected by default.

Fig. 2: General Tab

In the Y/ Dependent variables field, use your mouse to select the variable PRICE (see the tutorial on Selecting data for more information on this topic).

The prices are the "Ys" of the model as we want to predict these prices as a linear function of the other car attributes.

In X/ Predictors field, select the other 6 car attributes.

The name of the car models (MODEL) has also been selected as Observation labels.

To obtain the OLS regression solution, fix the number of components at 6, so it equals the number of predictors.  To accomplish this, in the Options tab set Number of components to ‘6’ and uncheck ‘Automatic’.

In the Options tab of the dialog box, make sure that the settings are as shown below.

Fig. 3: Options Tab

The fast computations start when you click on OK.

Following the basic statistics output section, the coefficients (unstandardized and standardized) are presented. For example, Table 3A presents the unstandardized coefficients.  Comparing Table 3A to Table 1, we see that the results match the OLS regression coefficients. 

Table 3A. Unstandardized coefficient estimates obtained from the 6-component (saturated) CCR model

These coefficients can be decomposed into parts associated with each of the 6 components using the component weights provided in Table 3B and the component coefficients (loadings) provided in Table 3C.

Table 3B. Unstandardized component weights 

Table 3C. Unstandardized loadings

For example, the coefficient     -1.94 for CYLINDER, can be decomposed as follows:

-1.94 = .006*(92.774) + .124*(1.381) + .804*(-3.728) + .627*(-11.016) + .422*(15.190) + .167*(5.053)

Re-open the CCR dialog box by selecting the Modeling data / Correlated Component Regression command in the Excel menu or click the corresponding button on the Modeling data toolbar.

Since N is relatively small (N=24) and the correlation between the predictors is fairly high, this saturated regression model overfits these data. We will now show how to activate the M-fold cross-validation (CV) option and show that this model is overfit, and that eliminating CCR components 3-6 provides the proper amount of regularization to produce more reliable results. To allow CV to assess all possible degrees of regularization, we will estimate all 6 CCR models (K≤6). We do this by activating the Automatic option in the Options tab.

The number of folds M is generally taken to be between 5 and 10, so we select M=6, the only integer between 5 and 10 that divides evenly into 24. In the Validation tab we activate ‘Cross-validation’ and request 100 rounds of 6-folds. By requesting more than 1 round, we obtain a standard error for the CV-R².  

Fig. 4: Validation Tab

Note that activating the ‘Automatic’ option also requests the Cross-Validation Component Plot to be generated (this is checked in the Charts tab) shown earlier in Fig. 1.

Click OK to perform these analyses. The Goodness of Fit Statistics show that the resulting model has K=2 components. For this model, the CV-R² increases to .750 with a standard error of only .014, providing a significant improvement over the OLS regression CV-R² =.64.

 Table 4A. Coefficients obtained from the 2-component model

 Table 4B. Component weights obtained from the 2-component model

Table 4C. Loadings obtained from the 2-component model

From the Coefficients Output in Tables 4A, 4B and 4C we see how the coefficients are now constructed based on only 2 components. For example, the coefficient for CYLINDER can be decomposed as follows:

20.944 = .221*92.774 + .349*1.381

Re-open the CCR dialog box by selecting the Modeling data / Correlated Component Regression command in the Excel menu or click the corresponding button on the Modeling data toolbar.

To eliminate extraneous and weak predictors, in the options tab we will now activate the step-down algorithm as shown below:

Figure 5. Options Tab

Activation of the step-down option automatically requests the step-down predictor selection plot in the Charts tab and the Predictor Count table from the Output tab.

Click on OK to estimate.

The predictor selection plot suggests that inclusion of 3 predictors in the model is optimal.

Figure 6. Cross-validation Step-down Plot

The Cross-validation Predictor Count table suggests that POWER and WEIGHT are the most important predictors, being included in 600 and 584 of the 1800 cross-validated regressions.

The final model has CV-R² = .77 and includes the predictors POWER, SPEED and WEIGHT:

Predictors retained in the model:

• POWER
• SPEED
• WEIGHT

General Discussion

Key driver regression attempts to ascertain the importance of several key explanatory variables (predictors) X₁, X₂, … , X_P that influence a dependent variable. For example, a typical dependent variable in key driver regression is “Customer Satisfaction”. Traditional OLS regression methods have difficulty with such derived importance tasks because the predictors usually have moderate to high correlation with each other, resulting in problems of confounding, making parameter estimates unstable and thus unusable as measures of importance.

Correlated Component Regression (CCR) is designed to handle such problems, and as shown in Tutorial 2 it even works with high-dimensional data where there are more predictors than cases!  Parameter estimates become more interpretable and cross-validation is used to avoid over-fitting, thus producing better out-of-sample predictions.

¿Dónde ir desde aquí? Volver al principio de la página