Abstract: In this study by Willerman et al.

(1991) the researchers use Magnetic Resonance Imaging (MRI) to determine the brain size of the subjects. They selected a sample of 40 right-handed Anglo introductory psychology students who had indicated no history of alcoholism, unconsciousness, brain damage, epilepsy, or heart disease. These subjects were drawn from a larger pool of introductory psychology students with total Scholastic Aptitude Test Scores higher than 1350 or lower than 940 who had agreed to satisfy a course requirement by allowing the administration of four subtests (Vocabulary, Similarities, Block Design, and Picture Completion) of the Wechsler (1981) Adult Intelligence Scale-Revised. With prior approval of the University's research review board, students selected for MRI were required to obtain prorated full-scale IQs of greater than 130 or less than 103. The MRI Scans were performed at the same facility for all 40 subjects. The scans consisted of 18 horizontal MR images. The computer counted all pixels with non-zero gray scale in each of the 18 images and the total count served as an index for brain size. Number of cases: 40 Variable Names: 1. FSIQ: Full Scale IQ scores based on the four Wechsler (1981) subtests 2. VIQ: Verbal IQ scores based on the four Wechsler (1981) subtests 3. PIQ: Performance IQ scores based on the four Wechsler (1981) subtests 4. MRI_Count: total pixel Count from the 18 MRI scans The Data:
816932 1001121 1038437 965353 951545 928799 991305 854258 904858 955466 833868 1079549 924059 856472 878897 865363 852244 945088 808020 889083 892420 135 140 137 133 137 99 138 92 89 133 132 138 134 140 96 83 132 100 101 80 83 132 150 126 129 132 90 136 90 93 114 129 150 129 120 100 71 132 96 112 77 83 125 125 150 128 134 110 133 96 84 147 121 129 126 147 90 96 120 102 84 86 86 905940 790619 955003 831772 935494 798612 1062462 793549 866662 857782 949589 997925 879987 834344 948066 949395 893983 930016 935863 816932 97 135 139 91 141 85 103 77 130 133 144 103 90 83 133 140 88 81 89 135 107 129 145 86 145 90 96 83 126 126 145 96 96 90 129 150 86 90 91 132 84 134 128 102 131 84 110 72 124 132 137 110 86 81 128 124 94 74 89 125

http://lib.stat.cmu.edu/DASL/Datafiles/Brainsize.html

Page 1 of 7

Output:
Regression Equation: MRI_Count = 620868.1 - 14449.2 * FSIQ + 8490.9 * VIQ + 8765.6 * PIQ Regression Statistics Multiple R 0.637403 R Square 0.406283 Adjusted R Square 0.356806 Standard Error 57969.74 Observations 40 ANOVA df Regression Residual Total SS MS 3 8.28E+10 2.76E+10 36 1.21E+11 3.36E+09 39 2.04E+11 F 8.21164 Significance F 0.000271442

Intercept FSIQ VIQ PIQ

Standard Coefficients Error 611455.00 62238.65 -17481.98 4516.58 10110.57 2624.86 10283.16 2474.47

t Stat 9.82 -3.87 3.85 4.16

P-value 0.00 0.00 0.00 0.00

Upper Lower 95% 95% 485229.16 737680.83 -26642.03 -8321.93 4787.10 15434.03 5264.71 15301.61

From the above summary output, the following can be observed and interpreted: 1. The coefficients of all the three parameters are statistically significant at 95% confidence level. This is further established by the p-value for these coefficients. 2. Adjusted R Square for the model suggests that close to 36% of the variations in the MRI_Count are explained by the variations in the aforementioned explanatory variable. As per Prof. Moradian, adjusted R square of more than 25% is meaningful in practical scenarios.

Page 2 of 7

Test for Multicollinearity:

Correlation Matrix MRI_Count FSIQ VIQ PIQ MRI_Count FSIQ VIQ 1 0.344984 1 0.342749 0.947743 1 0.395717 0.935121 0.786389 PIQ

Notice that all of our independent variables are highly correlated with one another. This suggests presence of multicollinearity in the model.

Remedy for Multicollinearity

A. Do nothing if the objective is to estimate y-values. B. Do a ratio transform. 1. After taking ratio of FSIQ, we get the following correlation matrix MRI_Count VIQ PIQ MRI_Count 1 VIQ 0.185743 1 PIQ 0.340036 0.79026 1 Here, PIQ and VIQ have strong correlation. 2. After taking ratio of VIQ, we get the following correlation matrix MRI_Count FSIQ PIQ MRI_Count 1 FSIQ 0.173132 1 PIQ 0.438647 0.939643 1 Here PIQ and FSIQ have strong correlation. 3. After taking ratio of PIQ, we get the following correlation matrix MRI_Count FSIQ MRI_Count 1 FSIQ 0.062086 1 VIQ 0.308097 0.949936 VIQ

Here VIQ and FSIQ have strong correlation. From the above transformations, it is evident that Multicollinearity is present even after taking the ratio transforms. To address Multicollinearity, one may add further data or may drop a variable. But dropping a variable may lead to specification bias and hence to Heteroscedasticity.

Page 3 of 7

Test for Heteroscedasticity: A. Graphical Method

Squared Residuals
2.5E+10 Residual Squared 2E+10 1.5E+10 1E+10 5E+09 0 0.00 500000.00 1000000.00 1500000.00 Predicted MRI_Count Squared Residuals

The above graph suggests that the Heteroscedasticity is not present in this model.
B. Parks Test: Regress LN(Squared Residuals) on each of the independent variable. 1. Regressing LN(Squared Residuals) on FSIQ, we do not get significant coefficients i.e. Squared R = 0.0004 p-value = 0.899896 2. Regressing LN(Squared Residuals) on VIQ, we do not get significant coefficients i.e. Squared R = 0.008 p-value = 0.577462 3. Regressing LN(Squared Residuals) on PIQ, we do not get significant coefficients i.e. Squared R = 0.004 p-value = 0.672219 Thus, From Parks test, we may conclude that Heteroscedasticity is not present. C. Whites Test: Regress squared residual on linear combination of the explanatory variables Squared R = 0.219 n*R^2 follows Chi-Square distribution with 10 degrees of freedom. Calculated Chi-Square = 8.764 Critical Chi-Square = 19.675 This is not significant. Hence, Heteroscedasticity is not present.

Page 4 of 7

Test for Autocorrelation:

A. Graphical Method: Scatter plot of Residuals on Residual(-1) shows

Residual(-1) (in 10000)

20.00 15.00 10.00 5.00 0.00 -20.00 -15.00 -10.00 -5.00 -5.00 -10.00 -15.00 -20.00 Residual(-1) 0.00 5.00 10.00 15.00 20.00

The above scatter plot suggests downward linear trends in the disturbances, hence, autocorrelation may be present. B. Durbin Watson d Test: Calculated d = 2.44 At 5% Level of Significance and k = 3, n=40 dL dU 4-dL 4-dU 1.338 1.659 2.662 2.341

As can be seen, the calculated d=2.44 lies in the range 4-dL and 4-dU. Hence, it lies in the zone of indecision. follows Standard Normal Distribution C. Here n = 40, (1-d/2) = -0.22 = 8.8 Thus, P(Z 8.8) 0 => Negative Autocorrelation exists

Page 5 of 7

Remedy for Autocorrelation:

A. Since, , , dont do anything. B. Using Generalized Least Square: Here, we require either true or From Durbin Watson d Test, we get From Cochrane Orkutt Method, we regress Residual on Residual(-1) and take the slope coefficient as Here, We will take the average of these estimates i.e. For GLS, we regress difference of MRI_Count and MRI_Count(-1) on the difference of explanatory variable and * explanatory variable. After regression, we test the model again for autocorrelation. We get the following output: 1. Graphical Method:

Residual(-1)
15 10 5 0 -20 -15 -10 -5 -5 -10 -15 -20 Residual(-1) 0 5 10 15

Slightly negative autocorrelation may be present. 2. Durbin Watson d Test: At 5% Level of Significance and k = 3, n=39 dL 1.338 dU 1.659 4-dL 2.662 4-dU 2.341 Calculated d = 2.13 lies in no autocorrelation region. follows Standard Normal Distribution 3. Here n = 39, (1-d/2) = -0.065 = 2.535 Thus, P(Z 2.535) is significant at 99% confidence level. Hence, no autocorrelation exists.

Page 6 of 7

Final result: The initial model had presence of Multicollinearity and autocorrelation, but Heteroscedasticity was absent. The model was treated for both Multicollinearity and autocorrelation and got rid of autocorrelation. Multicollinearity, though present, doesnt affect the outcome of the model. Summary output: Regression Statistics Multiple R 0.653357 R Square 0.426875 Adjusted R Square 0.37775 Standard Error 56298.65 Observations 39 ANOVA df Regression Residual Total 3 35 38 SS 8.26E+10 1.11E+11 1.94E+11 Significance MS F F 2.75E+10 8.689576 0.000192 3.17E+09

Intercept FSIQ- *FSIQ(-1) VIQ - *VIQ(-1) PIQ - *PIQ(-1)

Coefficients 773330.9 -17254.4 9689.362 10339

Standard Error t Stat P-value 70700.34 10.93815 7.71E-13 4394.338 -3.92652 0.000386 2514.888 3.852801 0.000477 2454.232 4.212724 0.000168

Lower 95% 629801.6 -26175.4 4583.869 5356.647

Upper 95% 916860.3 -8333.46 14794.86 15321.36

From the above summary output, the following can be observed and interpreted: 1. The coefficients of all the three transformed parameters are statistically significant at 95% confidence level. This is further established by the p-value for these coefficients. 2. Adjusted R Square for the model suggests that close to 38% of the variations in the MRI_Count are explained by the variations in the aforementioned explanatory variable. Regression equation: MRI_Count# = 620868.1 - 14449.2 * FSIQ# + 8490.9 * VIQ# + 8765.6 * PIQ# Here, # signify that the variable is rho-transformed i.e. GLS

Page 7 of 7