In this section, we are concerned with the confidence interval, called a "t-interval," for the mean response μ_Y when the predictor values are \(\textbf{X}_{h}=(1, X_{h,1}, X_{h,2}, \dots, X_{h,p-1})^\textrm{T}\). The general formula in words is as always:

Sample estimate ± (t-multiplier × standard error)

First, we define the standard error of the fit at \(\textbf{X}_{h}\) given by:

\(\textrm{se}(\hat{y}_{h})=\sqrt{\textrm{MSE}(\textbf{X}_{h}^{\textrm{T}}(\textbf{X}^{\textrm{T}}\textbf{X})^{-1}\textbf{X}_{h})}.\)

and the confidence interval is:

\(\hat{y}_h \pm t_{(1-\alpha/2, n-p)} \times \textrm{se}(\hat{y}_{h})\)

where:

• \(\hat{y}_h\) is the "fitted value" or "predicted value" of the response when the predictor values are \(\textbf{X}_{h}\).
• \(t_{(1-\alpha/2, n-p)}\) is the "t-multiplier." Note that the t-multiplier has n-p degrees of freedom because the confidence interval uses the mean square error (MSE) whose denominator is n-p.

Fortunately, we won't have to use the formula to calculate the confidence interval, since statistical software such as Minitab will do the dirty work for us. Here is some Minitab output for the example on IQ and physical characteristics from Lesson 5 (IQ Size data), where we've fit a model with PIQ as the response and Brain and Height as the predictors:

Here's what the output tells us:

• In the section labeled "Settings," Minitab reports the values for \(\textbf{X}_{h}\) (brain size = 90 and height = 70) for which we requested the confidence interval for \(\mu_{Y}\).
• In the section labeled "Prediction," Minitab reports a 95% confidence interval. We can be 95% confident that the average performance IQ score of all college students with brain size = 90 and height = 70 is between 98.24 and 113.04.
• In the section labeled "Prediction," Minitab also reports the predicted value \(\hat{y}_h\), ("Fit" = 105.64), the standard error of the fit ("SE Fit" = 3.65), and the 95% prediction interval for a new response (which we discuss in the next section).

Factors affecting the width of the t-interval for the mean response \(\left(\mu_{Y}\right)\)

As always, the formula is useful for investigating what factors affect the width of the confidence interval for \(\mu_{Y}\).

• As the mean square error (MSE) decreases, the width of the interval decreases. Since MSE is an estimate of how much the data vary naturally around the unknown population regression hyperplane, we have little control over MSE other than making sure that we make our measurements as carefully as possible.
• As we decrease the confidence level, the t-multiplier decreases, and hence the width of the interval decreases. In practice, we wouldn't want to set the confidence level below 90%.
• As we increase the sample size n, the width of the interval decreases. We have complete control over the size of our sample — the only limitation being our time and financial constraints.
• The closer \(\textbf{X}_{h}\) is to the average of the sample's predictor values, the narrower the interval.

Let's see this last claim in action for our IQ example:

The width of the first confidence interval we calculated earlier (113.04 - 98.24 = 14.80) is shorter than the width of this new interval (118.20 - 91.42 = 26.78), because 90 and 70 are much closer than 79 and 62 are to the sample means (90.7 and 68.4).

When is it okay to use the formula for the confidence interval for \(\mu_{Y}\) ?

• When \(\textbf{X}_{h}\) is within the "scope of the model." But, note that \(\textbf{X}_{h}\) does not have to be an actual observation in the data set.
• When the "LINE" conditions — linearity, independent errors, normal errors, equal error variances — are met. The formula works okay even if the error terms are only approximately normal. And, if you have a large sample, the error terms can even deviate substantially from normality.

In this section, we are concerned with the prediction interval for a new response \(y_{new}\) when the predictor values are  \(\textbf{X}_{h}=(1, X_{h,1}, X_{h,2}, \dots, X_{h,p-1})^\textrm{T}\). Again, let's just jump right in and learn the formula for the prediction interval. The general formula in words is as always:

and the formula in notation is:

\(\hat{y}_h \pm t_{(1-\alpha/2, n-p)} \times \sqrt{MSE + (\textrm{se}(\hat{y}_{h}))^2}\)

where:

• \(\hat{y}_h\) is the "fitted value" or "predicted value" of the response when the predictor values are \(\textbf{X}_{h}\).
• \(t_{(1-\alpha/2, n-p)}\) is the "t-multiplier." Note again that the t-multiplier has n-p degrees of freedom because the prediction interval uses the mean square error (MSE) whose denominator is n-p.
• \(\sqrt{MSE + (\textrm{se}(\hat{y}_{h}))^2}\) is the "standard error of the prediction," which is very similar to the "standard error of the fit" when estimating \(\mu_{Y}\). The standard error of the prediction just has an extra MSE term added that the standard error of the fit does not. (More on this a bit later.)

Again, we won't use the formula to calculate our prediction intervals. We'll let statistical software such as Minitab do the calculation for us. Let's look at the prediction interval for our IQ example (IQ Size data):

The output reports the 95% prediction interval for an individual college student with brain size = 90 and height = 70. We can be 95% confident that the performance IQ score of an individual college student with brain size = 90 and height = 70 will be between 65.35 and 145.93.

When is it okay to use the prediction interval for \(y_{new}\) formula?

The requirements are similar to, but a little more restrictive than, those for the confidence interval. It is okay:

• When \(\textbf{X}_{h}\) is within the "scope of the model." Again, \(\textbf{X}_{h}\) does not have to be an actual observation in the data set.
• When the "LINE" conditions — linearity, independent errors, normal errors, equal error variances — are met. Unlike the case for the formula for the confidence interval, the formula for the prediction interval depends strongly on the condition that the error terms are normally distributed.

Understanding the difference between the two formulas

In our discussion of the confidence interval for \(\mu_{Y}\), we used the formula to investigate what factors affect the width of the confidence interval. There's no need to do it again. Because the formulas are so similar, it turns out that the factors affecting the width of the prediction interval are identical to the factors affecting the width of the confidence interval.

Let's instead investigate the formula for the prediction interval for \(y_{new}\):

to see how it compares to the formula for the confidence interval for \(\mu_{Y}\):

Observe that the only difference in the formulas is that the standard error of the prediction for \(y_{new}\) has an extra MSE term in it that the standard error of the fit for \(\mu_{Y}\) does not. If you're not sure why this makes sense, re-read Section 3.3 on "Prediction Interval for a New Response" in the context of simple linear regression.

What are the practical implications of the difference between the two formulas?

• Because the prediction interval has the extra MSE term, a \(\left(1-\alpha \right)100\%\) confidence interval for \(\mu_{Y}\) at \(\textbf{X}_{h}\) will always be narrower than the corresponding (1-α)100% prediction interval for \(y_{new}\) at \(\textbf{X}_{h}\).
• By calculating the interval at the sample means of the predictor values and increasing the sample size n, the confidence interval's standard error can approach 0. Because the prediction interval has the extra MSE term, the prediction interval's standard error cannot get close to 0.

The four conditions ("LINE") that comprise the multiple linear regression model generalize the simple linear regression model conditions to take account of the fact that we now have multiple predictors:

• Linear Function: The mean of the response, \(\mbox{E}(Y_i)\), at each set of values of the predictors, \((x_{1i}, x_{2i},...)\), is a Linear function of the predictors.
• Independent: The errors, \( \epsilon_{i}\), are Independent.
• Normally Distributed: The errors, \( \epsilon_{i}\), at each set of values of the predictors, \((x_{1i}, x_{2i},...)\), are Normally distributed.
• Equal variances (denoted \(\alpha^{2}\)): The errors, \( \epsilon_{i}\), at each set of values of the predictors, \((x_{1i}, x_{2i},...)\), have Equal variances (denoted \(\alpha^{2}\)).

An equivalent way to think of the first (linearity) condition is that the mean of the error, \(\epsilon_i\), at each set of values of the predictors, \((x_{1i},x_{2i},\dots)\), is zero. An alternative way to describe all four assumptions is that the errors, \(\epsilon_i\), are independent normal random variables with mean zero and constant variance, \(\sigma^2\).

As in simple linear regression, we can assess whether these conditions seem to hold for a multiple linear regression model applied to a particular sample dataset by looking at the estimated errors, i.e., the residuals, \(e_i = y_i-\hat{y}_i\).

We can use all the methods we learned about in Lesson 4 to assess the multiple linear regression model assumptions:

• Create a scatterplot with the residuals, \(e_i\), on the vertical axis and the fitted values, \(\hat{y}_i\), on the horizontal axis and visual assess whether:
  • the (vertical) average of the residuals remains close to 0 as we scan the plot from left to right (this affirms the "L" condition);
  • the (vertical) spread of the residuals remains approximately constant as we scan the plot from left to right (this affirms the "E" condition);
  • there are no excessively outlying points (we'll explore this in more detail in Lesson 11).
  • violation of any of these three may necessitate remedial action (such as transforming one or more predictors and/or the response variable), depending on the severity of the violation (we'll explore this in more detail in Lesson 9).
• If the data observations were collected over time (or space) create a scatterplot with the residuals, \(e_i\), on the vertical axis and the time (or space) sequence on the horizontal axis and visual assess whether there is no systematic non-random pattern (this affirms the "I" condition).
  • Violation may suggest the need for a time series model (see the optional content).
• Create a series of scatterplots with the residuals, \(e_i\), on the vertical axis and each of the predictors in the model on the horizontal axes and visually assess whether:
  • the (vertical) average of the residuals remains close to 0 as we scan the plot from left to right (this affirms the "L" condition);
  • the (vertical) spread of the residuals remains approximately constant as we scan the plot from left to right (this affirms the "E" condition);
  • violation of either of these for at least one residual plot may suggest the need for transformations of one or more predictors and/or the response variable (again we'll explore this in more detail in Lesson 9).
• Create a histogram, boxplot, and/or normal probability plot of the residuals, \(e_i\) to check for approximate normality (the "N" condition). (Of these plots, the normal probability plot is generally the most effective.)
• Create a series of scatterplots with the residuals, \(e_i\), on the vertical axis and each of the available predictors that have been omitted from the model on the horizontal axes and visually assess whether:
  • there are no strong linear or simple nonlinear trends in the plot;
  • violation may indicate the predictor in question (or a transformation of the predictor) might be usefully added to the model.
  • it can sometimes be helpful to plot functions of predictor variables on the horizontal axis of a residual plot, for example, interaction terms consisting of one quantitative predictor multiplied by another quantitative predictor. A strong linear or simple nonlinear trend in the resulting plot may indicate the variable plotted on the horizontal axis might be usefully added to the model.

As you can see, checking the assumptions for a multiple linear regression model comprehensively is not a trivial undertaking! But, the more thorough we are in doing this, the greater the confidence we can have in our model. To illustrate, let's do a residual analysis for the example on IQ and physical characteristics from Lesson 5 (IQ Size data), where we've fit a model with PIQ as the response and Brain and Height as the predictors:

• First, here's a residual plot with the residuals, \(e_i\), on the vertical axis and the fitted values, \(\hat{y}_i\), on the horizontal axis:
  

  As we scan the plot from left to right, the average of the residuals remains approximately 0, the variation of the residuals appears to be roughly constant, and there are no excessively outlying points (except perhaps the observation with a residual of about 50, which might warrant some further investigation but isn't too much of a worry). Note that the two observations on the right of the plot with fitted values close to 140 are of no concern with respect to the model assumptions. We'd be reading too much into the plot if we were to worry that the residuals appear less variable on the right side of the plot (there are only 2 out of a total of 38 points here and hence there is little information on residual variability in this region of the plot).

• There is no time (or space) variable in this dataset so the next plot we'll consider is a scatterplot with the residuals, \(e_i\), on the vertical axis and one of the predictors in the model, Brain, on the horizontal axis:
  

  Again, as we scan the plot from left to right, the average of the residuals remains approximately 0, the variation of the residuals appears to be roughly constant, and there are no excessively outlying points. Also, there is no strong nonlinear trend in this plot that might suggest a transformation of PIQ or Brain in this model.

• The next plot we'll consider is a scatterplot with the residuals, \(e_i\), on the vertical axis and the other predictor in the model, Height, on the horizontal axis:
  

  Again, as we scan the plot from left to right, the average of the residuals remains approximately 0, the variation of the residuals appears to be roughly constant, and there are no excessively outlying points. Also, there is no strong nonlinear trend in this plot that might suggest a transformation of PIQ or Height in this model.

• The next plot we'll consider is a histogram of the residuals:
  

  Although this doesn't have the ideal bell-shaped appearance, given the small sample size there's little to suggest a violation of the normality assumption.

• Since the appearance of a histogram can be strongly influenced by the choice of intervals for the bars, to confirm this we can also look at a normal probability plot of the residuals:
  

  Again, given the small sample size, there's little to suggest a violation of the normality assumption.

• The final plot we'll consider for this example is a scatterplot with the residuals, \(e_i\), on the vertical axis and the only predictor excluded from the model, Weight, on the horizontal axis:
  

  Since there is no strong linear or simple nonlinear trend in this plot, there is nothing to suggest that Weight might be usefully added to the model. Don't get carried away by the apparent "up-down-up-down-up" pattern in this plot. This "trend" isn't nearly strong enough to warrant adding some complex function of Weight to the model - remember we've only got a sample size of 38 and we'd have to use up at least 5 degrees of freedom trying to add a fifth-degree polynomial of Weight to the model. All we'd end up doing if we did this is over-fitting the sample data and ending up with an over-complicated model that predicts new observations very poorly.

One key idea to draw from this example is that if you stare at a scatterplot of completely random points long enough you'll start to see patterns even when there are none! Residual analysis should be done thoroughly and carefully but without over-interpreting every slight anomaly. Serious problems with the multiple linear regression model generally reveal themselves pretty clearly in one or more residual plots. If a residual plot looks "mostly OK," chances are it is fine.

To complement the graphical methods just considered for assessing residual normality, we can perform a hypothesis test in which the null hypothesis is that the errors have a normal distribution. A large p-value and hence failure to reject this null hypothesis is a good result. It means that it is reasonable to assume that the errors have a normal distribution. Typically, assessment of the appropriate residual plots is sufficient to diagnose deviations from normality. However, a more rigorous and formal quantification of normality may be requested. So this section provides a discussion of some common testing procedures (of which there are many) for normality. For each test discussed below, the formal hypothesis test is written as:

\(\begin{align*} \nonumber H_{0}&\colon \textrm{the errors follow a normal distribution} \\ \nonumber H_{A}&\colon \textrm{the errors do not follow a normal distribution}. \end{align*}\)

While hypothesis tests are usually constructed to reject the null hypothesis, this is a case where we actually hope we fail to reject the null hypothesis as this would mean that the errors follow a normal distribution.

The Anderson-Darling Test measures the area between a fitted line (based on the chosen distribution) and a nonparametric step function (based on the plot points). The statistic is a squared distance that is weighted more heavily in the tails of the distribution. Smaller Anderson-Darling values indicate that the distribution fits the data better. The test statistic is given by:

\(\begin{equation*} A^{2}=-n-\sum_{i=1}^{n}\frac{2i-1}{n}(\log \textrm{F}(e_{i})+\log (1-\textrm{F}(e_{n+1-i}))), \end{equation*}\)

where \(\textrm{F}(\cdot)\) is the cumulative distribution of the normal distribution. The test statistic is compared against the critical values from a normal distribution in order to determine the p-value.

The Anderson-Darling test is available in some statistical software. To illustrate here's statistical software output for the example on IQ and physical characteristics from Lesson 5 (IQ Size data), where we've fit a model with PIQ as the response and Brain and Height as the predictors:

Since the Anderson-Darling test statistic is 0.262 with an associated p-value of 0.686, we fail to reject the null hypothesis and conclude that it is reasonable to assume that the errors have a normal distribution

The Shapiro-Wilk Test uses the test statistic

\(\begin{equation*} W=\dfrac{\biggl(\sum_{i=1}^{n}a_{i}e_{(i)}\biggr)^{2}}{\sum_{i=1}^{n}(e_{i}-\bar{e})^{2}}, \end{equation*} \)

where \(e_{i}\) pertains to the \(i^{th}\) largest value of the error terms and the \(a_i\) values are calculated using the means, variances, and covariances of the \(e_{i}\). W is compared against tabulated values of this statistic's distribution. Small values of W will lead to the rejection of the null hypothesis.

The Shapiro-Wilk test is available in some statistical software. For the IQ and physical characteristics model with PIQ as the response and Brain and Height as the predictors, the value of the test statistic is 0.976 with an associated p-value of 0.576, which leads to the same conclusion as for the Anderson-Darling test.

The Ryan-Joiner Test is a simpler alternative to the Shapiro-Wilk test. The test statistic is actually a correlation coefficient calculated by

\(\begin{equation*} R_{p}=\dfrac{\sum_{i=1}^{n}e_{(i)}z_{(i)}}{\sqrt{s^{2}(n-1)\sum_{i=1}^{n}z_{(i)}^2}}, \end{equation*}\)

where the \(z_{(i)}\) values are the z-score values (i.e., normal values) of the corresponding \(e_{(i)}\) value and \(s^{2}\) is the sample variance. Values of \(R_{p}\) closer to 1 indicate that the errors are normally distributed.

The Ryan-Joiner test is available in some statistical software. For the IQ and physical characteristics model with PIQ as the response and Brain and Height as the predictors, the value of the test statistic is 0.988 with an associated p-value > 0.1, which leads to the same conclusion as for the Anderson-Darling test.

The Kolmogorov-Smirnov Test (also known as the Lilliefors Test) compares the empirical cumulative distribution function of sample data with the distribution expected if the data were normal. If this observed difference is sufficiently large, the test will reject the null hypothesis of population normality. The test statistic is given by:

\(\begin{equation*} D=\max(D^{+},D^{-}), \end{equation*}\)

where

\(\begin{align*} D^{+}&=\max_{i}(i/n-\textrm{F}(e_{(i)}))\\ D^{-}&=\max_{i}(\textrm{F}(e_{(i)})-(i-1)/n), \end{align*}\)

The test statistic is compared against the critical values from a normal distribution in order to determine the p-value.

The Kolmogorov-Smirnov test is available in some statistical software. For the IQ and physical characteristics model with PIQ as the response and Brain and Height as the predictors, the value of the test statistic is 0.097 with an associated p-value of 0.490, which leads to the same conclusion as for the Anderson-Darling test.

There are various tests that may be performed on the residuals for testing if the regression errors have constant variance. It is usually sufficient to "visually" interpret residuals versus fitted values plots. However, the tests we discuss can provide an added layer of justification to your analysis. Note that some of the following procedures require you to partition the residuals into a certain number of groups, say \(g\geq 2\) groups of sizes \(n_{1},\ldots,n_{g}\) such that \(\sum_{i=1}^{g}n_{i}=n\). For these procedures, the sample variance of group i is given by:

\(\begin{equation*} s_{i}^{2}=\dfrac{\sum_{j=1}^{n_{i}}(e_{i,j}-\bar{e}_{i,\cdot})^{2}}{n_{i}-1} \end{equation*} \)

where \(e_{i,j}\) is the \(j^{\textrm{th}}\) residual from group i. Moreover, the pooled variance is given by:

\(\begin{equation*} s_{p}^{2}=\dfrac{\sum_{i=1}^{g}(n_{i}-1)s_{i}^{2}}{n-g} \end{equation*}\)

Suppose we partition the residuals of observations into two groups - one consisting of residuals associated with the lowest predictor values and the other consisting of those belonging to the highest predictor values. Treating these two groups as if they could (potentially) represent two different populations, we can test

\(\begin{align*} \nonumber H_{0}&\colon \sigma_{1}^{2}=\sigma_{2}^{2} \\ \nonumber H_{A}&\colon \sigma_{1}^{2}\neq\sigma_{2}^{2} \end{align*}\)

using the F-statistic \(F^{*}=s_{1}^{2}/s_{2}^{2}\). This test statistic is distributed according to a \(F_{n_{1}-1,n_{2}-1}\) distribution, so if \(F^{*}\geq F_{n_{1}-1,n_{2}-1;1-\alpha}\), then reject the null hypothesis and conclude that there is statistically significant evidence that the variance is not constant.

Another test for nonconstant variance is the modified Levene test (sometimes called the Brown-Forsythe test). This test does not require the error terms to be drawn from a normal distribution and hence it is a nonparametric test. The test is constructed by grouping the residuals into g groups according to the values of the quantity on the horizontal axis of the residual plot. It is typically recommended that each group has at least 25 observations and usually g=2 groups are used.

Begin by letting group 1 consist of the residuals associated with the \(n_{1}\) lowest values of the predictor. Then, let group 2 consists of the residuals associated with the \(n_{2}\) highest values of the predictor (so \(n_{1}+n_{2}=n\)). The objective is to perform the following hypothesis test:

\(\begin{align*} \nonumber H_{0}&\colon \textrm{the variance is constant} \\ \nonumber H_{A}&\colon \textrm{the variance is not constant} \end{align*}\)

As with the normality tests considered in the previous section, we hope we fail to reject the null hypothesis as this would mean the variance is constant. The test statistic for the above is computed as follows:

• \(d_{i,j}=|e_{i,j}-\tilde{e}_{i,\cdot}|\), where \(\tilde{e}_{i,\cdot}\) is the median of the \(i^{\textrm{th}}\) group of residuals
• \(s_{L}=\sqrt{\dfrac{\sum_{j=1}^{n_{1}}(d_{1,j}-\bar{d}_{1})^{2}+\sum_{j=1}^{n_{2}}(d_{2,j}-\bar{d}_{2})^{2}}{n_{1}+n_{2}-2}}\)
• \(L=\dfrac{\bar{d}_{1}-\bar{d}_{2}}{s_{L}\sqrt{\frac{1}{n_{1}}+\frac{1}{n_{2}}}}\)

L is approximately distributed according to a \(t_{n_{1}+n_{2}-2}\) distribution, or (equivalently) \(L^2\) is approximately distributed according to an \(F_{1,n_{1}+n_{2}-2}\) distribution.

To illustrate consider the Toluca Company dataset described on page 19 of Applied Linear Regression Models (4th ed) by Kutner, Nachtsheim, and Neter. We fit a simple linear regression model with predictor variable LotSize (the number of refrigerator parts in a production run) and response variable WorkHours (labor hours required to produce a lot of refrigerator parts). A modified Levene test applied with the 13 smallest lot sizes in the first group and the remaining 12 lot sizes in the second group results in the following:

• \(\tilde{e}_1=-19.87596\) and \(\tilde{e}_2=-2.68404\)
• \(\bar{d}_1=44.81507\) and \(\bar{d}_2=28.45034\)
• \(\sum{(d_1-\bar{d}_1)^2}=12566.61\) and \(\sum{(d_2-\bar{d}_2)^2}=9610.287\)
• \(s_L = \sqrt{(12566.61+9610.287)/23} = 31.05178\)
• \(L = (44.81507-28.40534)/(31.05178\sqrt{(1/13+1/12)}) = 1.31648\) and \(L^2=1.7331\)
• The \(t_{23}\) probability area to the left of 1.31648 is 0.8995, which means the p-value for the test is \(2(1-0.8995) = 0.201\), i.e., there is no evidence the errors have nonconstant variance.

The Breusch-Pagan test (also known as the Cook-Weisberg score test) is an alternative to the modified Levene test. Whereas the modified Levene test is a nonparametric test, the Breusch-Pagan test assumes that the error terms are normally distributed, with \(\mbox{E}(\epsilon_{i})=0\) and \(\mbox{Var}(\epsilon_{i})=\sigma^{2}_{i}\) (i.e., nonconstant variance). The \(\sigma_{i}^{2}\) values depend on the horizontal axis quantity (\(X_{i}\)) values in the following way:

\(\begin{equation*} \log\sigma_{i}^{2}=\gamma_{0}+\gamma_{1}X_{i} \end{equation*}\)

We are interested in testing the null hypothesis of constant variance versus the alternative hypothesis of nonconstant variance. Specifically, the hypothesis test is formulated as:

\(\begin{align*} \nonumber H_{0}&\colon \gamma_{1}=0 \\ \nonumber H_{A}&\colon \gamma_{1}\neq 0 \end{align*}\)

This test is carried out by first regressing the squared residuals on the predictor (i.e., regressing \(e_{i}^2\) on \(X_{i}\)). The sum of squares resulting from this analysis is denoted by \(\textrm{SSR}^{*}\), which provides a measure of the dependency of the error term on the predictor. The test statistic is given by:

\(\begin{equation*} X^{2*}=\dfrac{\textrm{SSR}^{*}/2}{(\textrm{SSE}/n)^{2}} \end{equation*}\)

where SSE is from the regression analysis of the response on the predictor. The p-value for this test is found using a \(\chi^{2}\) distribution with 1 degree of freedom (written as \(\chi^{2}_{1}\)).

To illustrate consider the Toluca Company example from above:

• \(SSE=54825\)
• \(SSR^*=7896142\)
• \(X^{2*} = (7896142/2) / (54825/25)^2 = 0.8209\)
• The \(\chi^{2}_{1}\) probability area to the left of 0.8209 is 0.635, which means the p-value for the test is \(1-0.635 = 0.365\), i.e., there is no evidence the errors have nonconstant variance.

When the variance is a function of q > 1 predictor, SSR^* derives from a multiple regression of the squared residuals against these predictors, and there are q degrees of freedom for the \(\chi^{2}\) distribution.

Bartlett's test is another test that can be used to test for constant variance. The objective is to again perform the following hypothesis test:

\(\begin{align*}\nonumber H_{0}\colon \textrm{the variance is constant} \\ \nonumber H_{A}\colon \textrm{the variance is not constant}\end{align*} \)

Bartlett's test is highly sensitive to the normality assumption, so if the residuals do not appear normal (even after transformations), then this test should not be used. Instead, the Levene test is the alternative to Bartlett's test which is less sensitive to departures from normality.

This test is carried out similarly to the Levene test. Once you have partitioned the residuals into g groups, the following test statistic can be constructed:

\(\begin{equation*} B=\dfrac{(n-g)\ln s_{p}^{2}-\sum_{i=1}^{g}(n_{i}-1)\ln s_{i}^{2}}{1+\biggl(\dfrac{1}{3(g-1)}\biggl((\sum_{i=1}^{g}\dfrac{1}{n_{i}-1})-\dfrac{1}{n-g}\biggr)\biggr)} \end{equation*}\)

The test statistic B is distributed according to a \(\chi^{2}_{g-1}\) distribution, so if \(B\geq \chi^{2}_{g-1;1-\alpha}\), then reject the null hypothesis and conclude that there is statistically significant evidence that the variance is not constant.

Data transformations are used in regression to describe curvature and can sometimes be usefully employed to correct for violation of the assumptions in a multiple linear regression model (particularly the "L" and "E" assumptions). We'll return to this topic in more detail in Lesson 9.

The next two pages cover the   Minitab and    commands for the procedures in this lesson.

Below is a text file that contains the data used in this lesson:

iqsize.txt