Generalized Linear Models

What Are Generalized Linear Models?

Linear regression models describe a linear relationship between a response and one or more predictive terms. Many times, however, a nonlinear relationship exists. Nonlinear Regression describes general nonlinear models. A special class of nonlinear models, called generalized linear models, uses linear methods.

Recall that linear models have these characteristics:

• At each set of values for the predictors, the response has a normal distribution with mean μ.

• A coefficient vector b defines a linear combination Xb of the predictors X.

• The model is μ = Xb.

In generalized linear models, these characteristics are generalized as follows:

• At each set of values for the predictors, the response has a distribution that can be normal, binomial, Poisson, gamma, or inverse Gaussian, with parameters including a mean μ.

• A coefficient vector b defines a linear combination Xb of the predictors X.

• A link function f defines the model as f(μ) = Xb.

Prepare Data

To begin fitting a regression, put your data into a form that fitting functions expect. All regression techniques begin with input data in an array X and response data in a separate vector y, or input data in a table or dataset array tbl and response data as a column in tbl. Each row of the input data represents one observation. Each column represents one predictor (variable).

For a table or dataset array tbl, indicate the response variable with the 'ResponseVar' name-value pair:

mdl = fitglm(tbl,'ResponseVar','BloodPressure');

The response variable is the last column by default.

You can use numeric categorical predictors. A categorical predictor is one that takes values from a fixed set of possibilities.

• For a numeric array X, indicate the categorical predictors using the 'Categorical' name-value pair. For example, to indicate that predictors 2 and 3 out of six are categorical:

mdl = fitglm(X,y,'Categorical',[2,3]);
% or equivalently
mdl = fitglm(X,y,'Categorical',logical([0 1 1 0 0 0]));
• For a table or dataset array tbl, fitting functions assume that these data types are categorical:

• Logical vector

• Categorical vector

• Character array

• String array

If you want to indicate that a numeric predictor is categorical, use the 'Categorical' name-value pair.

Represent missing numeric data as NaN. To represent missing data for other data types, see Missing Group Values.

• For a 'binomial' model with data matrix X, the response y can be:

• Binary column vector — Each entry represents success (1) or failure (0).

• Two-column matrix of integers — The first column is the number of successes in each observation, the second column is the number of trials in that observation.

• For a 'binomial' model with table or dataset tbl:

• Use the ResponseVar name-value pair to specify the column of tbl that gives the number of successes in each observation.

• Use the BinomialSize name-value pair to specify the column of tbl that gives the number of trials in each observation.

Dataset Array for Input and Response Data

For example, to create a dataset array from an Excel® spreadsheet:

ds = dataset('XLSFile','hospital.xls',...

To create a dataset array from workspace variables:

ds = dataset(MPG,Weight);
ds.Year = ordinal(Model_Year);

Table for Input and Response Data

To create a table from workspace variables:

tbl = table(MPG,Weight);
tbl.Year = ordinal(Model_Year);

Numeric Matrix for Input Data, Numeric Vector for Response

For example, to create numeric arrays from workspace variables:

X = [Weight Horsepower Cylinders Model_Year];
y = MPG;

To create numeric arrays from an Excel spreadsheet:

y = X(:,4); % response y is systolic pressure
X(:,4) = []; % remove y from the X matrix

Notice that the nonnumeric entries, such as sex, do not appear in X.

Choose Generalized Linear Model and Link Function

Often, your data suggests the distribution type of the generalized linear model.

Response Data TypeSuggested Model Distribution Type
Any real number'normal'
Any positive number'gamma' or 'inverse gaussian'
Any nonnegative integer'poisson'
Integer from 0 to n, where n is a fixed positive value'binomial'

Set the model distribution type with the Distribution name-value pair. After selecting your model type, choose a link function to map between the mean µ and the linear predictor Xb.

ValueDescription
'comploglog'

log(–log((1 – µ))) = Xb

'identity', default for the distribution 'normal'

µ = Xb

'log', default for the distribution 'poisson'

log(µ) = Xb

'logit', default for the distribution 'binomial'

log(µ/(1 – µ)) = Xb

'loglog'

log(–log(µ)) = Xb

'probit'

Φ–1(µ) = Xb, where Φ is the normal (Gaussian) cumulative distribution function

'reciprocal', default for the distribution 'gamma'

µ–1 = Xb

p (a number), default for the distribution 'inverse gaussian' (with p = –2)

µp = Xb

A cell array of the form {FL FD FI}, containing three function handles created using @, which define the link (FL), the derivative of the link (FD), and the inverse link (FI). Or, a structure of function handles with the field Link containing FL, the field Derivative containing FD, and the field Inverse containing FI.

The nondefault link functions are mainly useful for binomial models. These nondefault link functions are 'comploglog', 'loglog', and 'probit'.

The link function defines the relationship f(µ) = Xb between the mean response µ and the linear combination Xb = X*b of the predictors. You can choose one of the built-in link functions or define your own by specifying the link function FL, its derivative FD, and its inverse FI:

• The link function FL calculates f(µ).

• The derivative of the link function FD calculates df(µ)/.

• The inverse function FI calculates g(Xb) = µ.

You can specify a custom link function in either of two equivalent ways. Each way contains function handles that accept a single array of values representing µ or Xb, and returns an array the same size. The function handles are either in a cell array or a structure:

• Cell array of the form {FL FD FI}, containing three function handles, created using @, that define the link (FL), the derivative of the link (FD), and the inverse link (FI).

• Structure s with three fields, each containing a function handle created using @:

• s.Derivative — Derivative of the link function

• s.Inverse — Inverse of the link function

For example, to fit a model using the 'probit' link function:

x = [2100 2300 2500 2700 2900 ...
3100 3300 3500 3700 3900 4100 4300]';
n = [48 42 31 34 31 21 23 23 21 16 17 21]';
y = [1 2 0 3 8 8 14 17 19 15 17 21]';
g = fitglm(x,[y n],...
g =

Generalized Linear regression model:
probit(y) ~ 1 + x1
Distribution = Binomial

Estimated Coefficients:
Estimate     SE            tStat     pValue
(Intercept)      -7.3628       0.66815    -11.02    3.0701e-28
x1             0.0023039    0.00021352     10.79    3.8274e-27

12 observations, 10 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 241, p-value = 2.25e-54

You can perform the same fit using a custom link function that performs identically to the 'probit' link function:

s = {@norminv,@(x)1./normpdf(norminv(x)),@normcdf};
g = fitglm(x,[y n],...
g =

Generalized Linear regression model:
Distribution = Binomial

Estimated Coefficients:
Estimate     SE            tStat     pValue
(Intercept)      -7.3628       0.66815    -11.02    3.0701e-28
x1             0.0023039    0.00021352     10.79    3.8274e-27

12 observations, 10 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 241, p-value = 2.25e-54

The two models are the same.

Equivalently, you can write s as a structure instead of a cell array of function handles:

s.Derivative = @(x) 1./normpdf(norminv(x));
s.Inverse = @normcdf;
g = fitglm(x,[y n],...
g =

Generalized Linear regression model:
Distribution = Binomial

Estimated Coefficients:
Estimate     SE            tStat     pValue
(Intercept)      -7.3628       0.66815    -11.02    3.0701e-28
x1             0.0023039    0.00021352     10.79    3.8274e-27

12 observations, 10 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 241, p-value = 2.25e-54

Choose Fitting Method and Model

There are two ways to create a fitted model.

• Use fitglm when you have a good idea of your generalized linear model, or when you want to adjust your model later to include or exclude certain terms.

• Use stepwiseglm when you want to fit your model using stepwise regression. stepwiseglm starts from one model, such as a constant, and adds or subtracts terms one at a time, choosing an optimal term each time in a greedy fashion, until it cannot improve further. Use stepwise fitting to find a good model, one that has only relevant terms.

The result depends on the starting model. Usually, starting with a constant model leads to a small model. Starting with more terms can lead to a more complex model, but one that has lower mean squared error.

In either case, provide a model to the fitting function (which is the starting model for stepwiseglm).

Specify a model using one of these methods.

Brief Model Name

NameModel Type
'constant'Model contains only a constant (intercept) term.
'linear'Model contains an intercept and linear terms for each predictor.
'interactions'Model contains an intercept, linear terms, and all products of pairs of distinct predictors (no squared terms).
'purequadratic'Model contains an intercept, linear terms, and squared terms.
'quadratic'Model contains an intercept, linear terms, interactions, and squared terms.
'polyijk'Model is a polynomial with all terms up to degree i in the first predictor, degree j in the second predictor, etc. Use numerals 0 through 9. For example, 'poly2111' has a constant plus all linear and product terms, and also contains terms with predictor 1 squared.

Terms Matrix

A terms matrix T is a t-by-(p + 1) matrix specifying terms in a model, where t is the number of terms, p is the number of predictor variables, and +1 accounts for the response variable. The value of T(i,j) is the exponent of variable j in term i.

For example, suppose that an input includes three predictor variables x1, x2, and x3 and the response variable y in the order x1, x2, x3, and y. Each row of T represents one term:

• [0 0 0 0] — Constant term or intercept

• [0 1 0 0]x2; equivalently, x1^0 * x2^1 * x3^0

• [1 0 1 0]x1*x3

• [2 0 0 0]x1^2

• [0 1 2 0]x2*(x3^2)

The 0 at the end of each term represents the response variable. In general, a column vector of zeros in a terms matrix represents the position of the response variable. If you have the predictor and response variables in a matrix and column vector, then you must include 0 for the response variable in the last column of each row.

Formula

A formula for a model specification is a character vector or string scalar of the form

'y ~ terms',

• y is the response name.

• terms contains

• Variable names

• + to include the next variable

• - to exclude the next variable

• : to define an interaction, a product of terms

• * to define an interaction and all lower-order terms

• ^ to raise the predictor to a power, exactly as in * repeated, so ^ includes lower order terms as well

• () to group terms

Tip

Formulas include a constant (intercept) term by default. To exclude a constant term from the model, include -1 in the formula.

Examples:

'y ~ x1 + x2 + x3' is a three-variable linear model with intercept.
'y ~ x1 + x2 + x3 - 1' is a three-variable linear model without intercept.
'y ~ x1 + x2 + x3 + x2^2' is a three-variable model with intercept and a x2^2 term.
'y ~ x1 + x2^2 + x3' is the same as the previous example, since x2^2 includes a x2 term.
'y ~ x1 + x2 + x3 + x1:x2' includes an x1*x2 term.
'y ~ x1*x2 + x3' is the same as the previous example, since x1*x2 = x1 + x2 + x1:x2.
'y ~ x1*x2*x3 - x1:x2:x3' has all interactions among x1, x2, and x3, except the three-way interaction.
'y ~ x1*(x2 + x3 + x4)' has all linear terms, plus products of x1 with each of the other variables.

Fit Model to Data

Create a fitted model using fitglm or stepwiseglm. Choose between them as in Choose Fitting Method and Model. For generalized linear models other than those with a normal distribution, give a Distribution name-value pair as in Choose Generalized Linear Model and Link Function. For example,

mdl = fitglm(X,y,'linear','Distribution','poisson')
% or
'Distribution','binomial')

Examine Quality and Adjust the Fitted Model

After fitting a model, examine the result.

Model Display

A linear regression model shows several diagnostics when you enter its name or enter disp(mdl). This display gives some of the basic information to check whether the fitted model represents the data adequately.

For example, fit a Poisson model to data constructed with two out of five predictors not affecting the response, and with no intercept term:

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[.4;.2;.3]);
y = poissrnd(mu);
mdl = fitglm(X,y,...
'linear','Distribution','poisson')
mdl =

Generalized Linear regression model:
log(y) ~ 1 + x1 + x2 + x3 + x4 + x5
Distribution = Poisson

Estimated Coefficients:
Estimate     SE          tStat       pValue
(Intercept)     0.039829     0.10793     0.36901       0.71212
x1               0.38551    0.076116      5.0647    4.0895e-07
x2             -0.034905    0.086685    -0.40266        0.6872
x3              -0.17826    0.093552     -1.9054      0.056722
x4               0.21929     0.09357      2.3436      0.019097
x5               0.28918      0.1094      2.6432     0.0082126

100 observations, 94 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 44.9, p-value = 1.55e-08

Notice that:

• The display contains the estimated values of each coefficient in the Estimate column. These values are reasonably near the true values [0;.4;0;0;.2;.3], except possibly the coefficient of x3 is not terribly near 0.

• There is a standard error column for the coefficient estimates.

• The reported pValue (which are derived from the t statistics under the assumption of normal errors) for predictors 1, 4, and 5 are small. These are the three predictors that were used to create the response data y.

• The pValue for (Intercept), x2 and x3 are larger than 0.01. These three predictors were not used to create the response data y. The pValue for x3 is just over .05, so might be regarded as possibly significant.

• The display contains the Chi-square statistic.

Diagnostic Plots

Diagnostic plots help you identify outliers, and see other problems in your model or fit. To illustrate these plots, consider binomial regression with a logistic link function.

The logistic model is useful for proportion data. It defines the relationship between the proportion p and the weight w by:

log[p/(1 – p)] = b1 + b2w

This example fits a binomial model to data. The data are derived from carbig.mat, which contains measurements of large cars of various weights. Each weight in w has a corresponding number of cars in total and a corresponding number of poor-mileage cars in poor.

It is reasonable to assume that the values of poor follow binomial distributions, with the number of trials given by total and the percentage of successes depending on w. This distribution can be accounted for in the context of a logistic model by using a generalized linear model with link function log(µ/(1 – µ)) = Xb. This link function is called 'logit'.

w = [2100 2300 2500 2700 2900 3100 ...
3300 3500 3700 3900 4100 4300]';
total = [48 42 31 34 31 21 23 23 21 16 17 21]';
poor = [1 2 0 3 8 8 14 17 19 15 17 21]';
mdl = fitglm(w,[poor total],...
mdl =

Generalized Linear regression model:
logit(y) ~ 1 + x1
Distribution = Binomial

Estimated Coefficients:
Estimate     SE            tStat      pValue
(Intercept)       -13.38         1.394    -9.5986    8.1019e-22
x1             0.0041812    0.00044258     9.4474    3.4739e-21

12 observations, 10 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 242, p-value = 1.3e-54

See how well the model fits the data.

plotSlice(mdl) The fit looks reasonably good, with fairly wide confidence bounds.

To examine further details, create a leverage plot.

plotDiagnostics(mdl) This is typical of a regression with points ordered by the predictor variable. The leverage of each point on the fit is higher for points with relatively extreme predictor values (in either direction) and low for points with average predictor values. In examples with multiple predictors and with points not ordered by predictor value, this plot can help you identify which observations have high leverage because they are outliers as measured by their predictor values.

Residuals — Model Quality for Training Data

There are several residual plots to help you discover errors, outliers, or correlations in the model or data. The simplest residual plots are the default histogram plot, which shows the range of the residuals and their frequencies, and the probability plot, which shows how the distribution of the residuals compares to a normal distribution with matched variance.

This example shows residual plots for a fitted Poisson model. The data construction has two out of five predictors not affecting the response, and no intercept term:

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[2;1;.5]);
y = poissrnd(mu);
mdl = fitglm(X,y,...
'linear','Distribution','poisson');

Examine the residuals:

plotResiduals(mdl) While most residuals cluster near 0, there are several near ±18. So examine a different residuals plot.

plotResiduals(mdl,'fitted') The large residuals don’t seem to have much to do with the sizes of the fitted values.

plotResiduals(mdl,'probability') Now it is clear. The residuals do not follow a normal distribution. Instead, they have fatter tails, much as an underlying Poisson distribution.

Plots to Understand Predictor Effects and How to Modify a Model

This example shows how to understand the effect each predictor has on a regression model, and how to modify the model to remove unnecessary terms.

1. Create a model from some predictors in artificial data. The data do not use the second and third columns in X. So you expect the model not to show much dependence on those predictors.

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[2;1;.5]);
y = poissrnd(mu);
mdl = fitglm(X,y,...
'linear','Distribution','poisson');
2. Examine a slice plot of the responses. This displays the effect of each predictor separately.

plotSlice(mdl) The scale of the first predictor is overwhelming the plot. Disable it using the Predictors menu.   Now it is clear that predictors 2 and 3 have little to no effect.

You can drag the individual predictor values, which are represented by dashed blue vertical lines. You can also choose between simultaneous and non-simultaneous confidence bounds, which are represented by dashed red curves. Dragging the predictor lines confirms that predictors 2 and 3 have little to no effect.

3. Remove the unnecessary predictors using either removeTerms or step. Using step can be safer, in case there is an unexpected importance to a term that becomes apparent after removing another term. However, sometimes removeTerms can be effective when step does not proceed. In this case, the two give identical results.

mdl1 = removeTerms(mdl,'x2 + x3')
mdl1 =

Generalized Linear regression model:
log(y) ~ 1 + x1 + x4 + x5
Distribution = Poisson

Estimated Coefficients:
Estimate    SE          tStat     pValue
(Intercept)    0.17604     0.062215    2.8295       0.004662
x1              1.9122     0.024638    77.614              0
x4             0.98521     0.026393    37.328    5.6696e-305
x5             0.61321     0.038435    15.955     2.6473e-57

100 observations, 96 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 4.97e+04, p-value = 0
mdl1 = step(mdl,'NSteps',5,'Upper','linear')
1. Removing x3, Deviance = 93.856, Chi2Stat = 0.00075551, PValue = 0.97807
2. Removing x2, Deviance = 96.333, Chi2Stat = 2.4769, PValue = 0.11553

mdl1 =

Generalized Linear regression model:
log(y) ~ 1 + x1 + x4 + x5
Distribution = Poisson

Estimated Coefficients:
Estimate    SE          tStat     pValue
(Intercept)    0.17604     0.062215    2.8295       0.004662
x1              1.9122     0.024638    77.614              0
x4             0.98521     0.026393    37.328    5.6696e-305
x5             0.61321     0.038435    15.955     2.6473e-57

100 observations, 96 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 4.97e+04, p-value = 0

Predict or Simulate Responses to New Data

There are three ways to use a linear model to predict the response to new data:

predict

The predict method gives a prediction of the mean responses and, if requested, confidence bounds.

This example shows how to predict and obtain confidence intervals on the predictions using the predict method.

1. Create a model from some predictors in artificial data. The data do not use the second and third columns in X. So you expect the model not to show much dependence on these predictors. Construct the model stepwise to include the relevant predictors automatically.

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[2;1;.5]);
y = poissrnd(mu);
mdl = stepwiseglm(X,y,...
'constant','upper','linear','Distribution','poisson');
1. Adding x1, Deviance = 2515.02869, Chi2Stat = 47242.9622, PValue = 0
2. Adding x4, Deviance = 328.39679, Chi2Stat = 2186.6319, PValue = 0
3. Adding x5, Deviance = 96.3326, Chi2Stat = 232.0642, PValue = 2.114384e-52
2. Generate some new data, and evaluate the predictions from the data.

Xnew = randn(3,5) + repmat([1 2 3 4 5],[3,1]); % new data
[ynew,ynewci] = predict(mdl,Xnew)
ynew =

1.0e+04 *

0.1130
1.7375
3.7471

ynewci =

1.0e+04 *

0.0821    0.1555
1.2167    2.4811
2.8419    4.9407

feval

When you construct a model from a table or dataset array, feval is often more convenient for predicting mean responses than predict. However, feval does not provide confidence bounds.

This example shows how to predict mean responses using the feval method.

1. Create a model from some predictors in artificial data. The data do not use the second and third columns in X. So you expect the model not to show much dependence on these predictors. Construct the model stepwise to include the relevant predictors automatically.

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[2;1;.5]);
y = poissrnd(mu);

X = array2table(X); % create data table
y = array2table(y);
tbl = [X y];

mdl = stepwiseglm(tbl,...
'constant','upper','linear','Distribution','poisson');
1. Adding x1, Deviance = 2515.02869, Chi2Stat = 47242.9622, PValue = 0
2. Adding x4, Deviance = 328.39679, Chi2Stat = 2186.6319, PValue = 0
3. Adding x5, Deviance = 96.3326, Chi2Stat = 232.0642, PValue = 2.114384e-52
2. Generate some new data, and evaluate the predictions from the data.

Xnew = randn(3,5) + repmat([1 2 3 4 5],[3,1]); % new data
ynew = feval(mdl,Xnew(:,1),Xnew(:,4),Xnew(:,5)) % only need predictors 1,4,5
ynew =

1.0e+04 *

0.1130
1.7375
3.7471

Equivalently,

ynew = feval(mdl,Xnew(:,[1 4 5])) % only need predictors 1,4,5
ynew =

1.0e+04 *

0.1130
1.7375
3.7471

random

The random method generates new random response values for specified predictor values. The distribution of the response values is the distribution used in the model. random calculates the mean of the distribution from the predictors, estimated coefficients, and link function. For distributions such as normal, the model also provides an estimate of the variance of the response. For the binomial and Poisson distributions, the variance of the response is determined by the mean; random does not use a separate “dispersion” estimate.

This example shows how to simulate responses using the random method.

1. Create a model from some predictors in artificial data. The data do not use the second and third columns in X. So you expect the model not to show much dependence on these predictors. Construct the model stepwise to include the relevant predictors automatically.

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[2;1;.5]);
y = poissrnd(mu);
mdl = stepwiseglm(X,y,...
'constant','upper','linear','Distribution','poisson');
1. Adding x1, Deviance = 2515.02869, Chi2Stat = 47242.9622, PValue = 0
2. Adding x4, Deviance = 328.39679, Chi2Stat = 2186.6319, PValue = 0
3. Adding x5, Deviance = 96.3326, Chi2Stat = 232.0642, PValue = 2.114384e-52
2. Generate some new data, and evaluate the predictions from the data.

Xnew = randn(3,5) + repmat([1 2 3 4 5],[3,1]); % new data
ysim = random(mdl,Xnew)
ysim =

1111
17121
37457

The predictions from random are Poisson samples, so are integers.

3. Evaluate the random method again, the result changes.

ysim = random(mdl,Xnew)
ysim =

1175
17320
37126

Share Fitted Models

The model display contains enough information to enable someone else to recreate the model in a theoretical sense. For example,

rng('default') % for reproducibility
X = randn(100,5);
mu = exp(X(:,[1 4 5])*[2;1;.5]);
y = poissrnd(mu);
mdl = stepwiseglm(X,y,...
'constant','upper','linear','Distribution','poisson')
1. Adding x1, Deviance = 2515.02869, Chi2Stat = 47242.9622, PValue = 0
2. Adding x4, Deviance = 328.39679, Chi2Stat = 2186.6319, PValue = 0
3. Adding x5, Deviance = 96.3326, Chi2Stat = 232.0642, PValue = 2.114384e-52

mdl =

Generalized Linear regression model:
log(y) ~ 1 + x1 + x4 + x5
Distribution = Poisson

Estimated Coefficients:
Estimate    SE          tStat     pValue
(Intercept)    0.17604     0.062215    2.8295       0.004662
x1              1.9122     0.024638    77.614              0
x4             0.98521     0.026393    37.328    5.6696e-305
x5             0.61321     0.038435    15.955     2.6473e-57

100 observations, 96 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 4.97e+04, p-value = 0

You can access the model description programmatically, too. For example,

mdl.Coefficients.Estimate
ans =

0.1760
1.9122
0.9852
0.6132
mdl.Formula
ans =

log(y) ~ 1 + x1 + x4 + x5

 Collett, D. Modeling Binary Data. New York: Chapman & Hall, 2002.

 Dobson, A. J. An Introduction to Generalized Linear Models. New York: Chapman & Hall, 1990.

 McCullagh, P., and J. A. Nelder. Generalized Linear Models. New York: Chapman & Hall, 1990.

 Neter, J., M. H. Kutner, C. J. Nachtsheim, and W. Wasserman. Applied Linear Statistical Models, Fourth Edition. Irwin, Chicago, 1996.