Generalized Linear Models

About the Generalized Linear Models Task

Example: Analyzing the Sashelp.Baseball Data Set

Assigning Data to Roles

Building a Model

Requirements for Building a Model

Create a Main Effect

Create Crossed Effects (Interactions)

Create a Nested Effect

Create a Full Factorial Model

Create N-Way Factorial

Create Polynomial Effects of the Nth Order

Specifying Model Effects for Zero-Inflated Models

Setting Options

Setting the Output Options

About the Generalized Linear Models Task

Generalized linear models are an extension of traditional linear models. In a generalized linear model, the mean of a population depends on a linear predictor through a nonlinear link function. The response probability distribution can be any member of the exponential family of distributions. Examples of generalized linear models include classical linear models with normal errors, logistic and probit models for binary data, and log-linear models for multinomial data. Other statistical models can be formulated as generalized linear models by the selection of an appropriate link function and response probability distribution.

The Generalized Linear Models task provides model fitting and model building for generalized linear models. It fits models for standard distributions such as Normal, Poisson, and Tweedie in the exponential family. This task also fits multinomial models for ordinal and nominal responses. The task provides forward, backward, and stepwise selection methods.

Example: Analyzing the Sashelp.Baseball Data Set

To create this example:

In the Tasks section, expand the Statistics folder and double-click Generalized Linear Models. The user interface for the Generalized Linear Models task opens.
On the Data tab, select the SASHELP.BASEBALL data set.

From the Distribution drop-down list, select Poisson. Assign columns to these roles:

Role	Column Name
Response
Response variable	nHome From the Link function drop-down list, select Logarithm.
Explanatory Variables
Classification variables	League
Continuous variables	logSalary

Click the Model tab. In the Variables box, select League and logSalary. Click Add to add these as main effects.
To run the task, click .

Here is a subset of the results:

Assigning Data to Roles

To run the Generalized Linear Models task, you must assign a column to the Response variable role for all distribution types except binomial. If you select a binomial distribution, you must assign either a single response variable or a pair of variables to the Number of events and Number of trials roles.

Option Name	Description
Roles
Response
Distribution	specifies the distribution for your model. You can choose from these distributions: Binomial Gamma Inverse Gaussian Multinomial Negative binomial Normal Poisson Tweedie. If you select a Tweedie distribution, you can specify the Tweedie power parameter. This value can be 0, 1, or a value greater than 1.1 and less than or equal to 3.0. Zero-inflated negative binomial Zero-inflated Poisson
Options for Binomial Distribution
Response data consists of numbers of events and trials	specifies that a pair of variables consists of response data for events and trials.
Number of events	specifies the column that contains the number of events.
Number of trials	specifies the column that contains the number of trials.
Response	specifies the single variable that contains response values. Use the Event of interest option to select a value of the response variable that represents the event that you want to model. Note: The Response role and the Event of interest option are available only if you do not select the Response data consists of numbers of events and trials check box.
Options for All Distribution Types
Response	specifies the variable that contains the response data. For most distribution types, you specify a single numeric variable.
Link function	specifies the link function for your model. The functions that are available depend on the selected distribution.
Explanatory Variables
Classification variables	specifies the variables to use to group (classify) data in the analysis. Classification variables can be either character or numeric. A classification variable is a variable that enters the statistical analysis or model through its levels, not through its values. The process of associating values of a variable with levels is termed levelization.
Parameterization of Effects
Coding	specifies the parameterization method for the classification variable. Design matrix columns are created from the classification variables according to the selected coding scheme. You can select from these coding schemes: Effect coding specifies effect coding. GLM coding specifies less-than-full-rank, reference-cell coding. This coding scheme is the default. Reference coding specifies reference-cell coding.
Treatment of Missing Values
An observation is excluded from the analysis when either of these conditions is met: if any variable in the model contains a missing value if any classification variable contains a missing value (regardless of whether the classification variable is used in the model)
Continuous variables	specifies the independent covariates (regressors) for the regression model. If you do not specify a continuous variable, the task fits a model that contains only an intercept.
Offset variable	specifies a variable to be used as an offset to the linear predictor. An offset plays the role of an effect whose coefficient is known to be 1. Observations that have missing values for the offset variable are excluded from the analysis.
Additional Roles
Frequency count	specifies the numeric column that contains the frequency of occurrence for each observation.
Weight variable	specifies the numeric column to use as a weight to perform a weighted analysis of the data.
Group analysis by	specifies the column to use as the BY variable.

Building a Model

Requirements for Building a Model

By default, no effects are specified, which results in the task fitting an intercept-only model. To specify an effect, you must assign at least one variable to the Classification variables role or the Continuous variables role. You can select combinations of variables to create crossed, nested, factorial, or polynomial effects.

To create a model, use the model builder on the Models tab. After you create the model, you can specify whether to include the intercept in the model.

Create a Main Effect

Select the variable name in the Variables box.
Click Add to add the variable to the Model effects box.

Create Crossed Effects (Interactions)

Select two or more variables in the Variables box. To select more than one variable, press Ctrl.
Click Cross.

Create a Nested Effect

Nested effects are specified by following a main effect or crossed effect with a classification variable or list of classification variables enclosed in parentheses. The main effect or crossed effect is nested within the effects listed in parentheses. Here are examples of nested effects: B(A), C(B*A), D*E(C*B*A). In this example, B(A) is read "A nested within B."

Select the effect name in the Model effects box.
Click Nest. The Nested window opens.
Select the variable to use in the nested effect. Click Outer or Nested within Outer to specify how to create the nested effect.

Note: The Nested within Outer button is available only when a classification variable is selected.
Click Add.

Create a Full Factorial Model

Select two or more variables in the Variables box.
Click Full Factorial.

For example, if you select the Height, Weight, and Age variables and then click Full Factorial, these model effects are created: Age, Height, Weight, Age*Height, Age*Weight, Height*Weight, and Age*Height*Weight.

Create N-Way Factorial

Select two or more variables in the Variables box.
Click N-way Factorial to add these effects to the Model effects box.

For example, if you select the Height, Weight, and Age variables and then specify the value of N as 2, when you click N-way Factorial, these model effects are created: Age, Height, Weight, Age*Height, Age*Weight, and Height*Weight. If N is set to a value greater than the number of variables in the model, N is effectively set to the number of variables.

Create Polynomial Effects of the Nth Order

Select one variable in the Variables box.
Specify higher-degree crossings by adjusting the number in the N field.
Click Polynomial Order=N to add the polynomial effects to the Model effects box.

For example, if you select the Age and Height variables and then you specify 3 in the N field, when you click Polynomial Order=N, these model effects are created: Age, Age*Age, Age*Age*Age, Height, Height*Height, and Height*Height*Height.

Specifying Model Effects for Zero-Inflated Models

These options are available if you selected Zero-inflated negative binomial or Zero-inflated Poisson as the distribution on the Data tab.

You must choose the type of model that you want to create:

an intercept-only model.
a model that includes effects from the main model. You define these model effects by using the model builder.
a custom model. You specify these effects in the Enter a custom model text box. If you specify multiple effects, use a space between each effect.

If you choose to include effects in the zero-inflated models, specify the link function for those effects.

Setting Options

Option	Description
Methods
Dispersion
Adjust for overdispersion	adjusts the parameter covariance matrix and the likelihood function by a scale parameter. For the dispersion parameter, you can select a Pearson estimate or a deviance estimate. To define the subpopulations for calculating the Pearson and deviance chi-square goodness-of-fit tests, assign one or more variables to the role. Note: This option is available only for binomial and multinomial distributions.
Estimate dispersion parameter	enables you to specify a fixed dispersion parameter for those distributions that have a dispersion parameter. By default, this parameter is estimated. Note: This option is not available for binomial and multinomial distributions, but it is available for the other distribution types.
Optimization
Maximum number of iterations	specifies the maximum number of iterations to perform for the selected optimization technique.
Statistics
You can select the statistics to include in the output. Here are the additional statistics that you can include: type 1 (sequential) analysis type 3 analysis Wald statistics for Type 3 contrasts confidence intervals, such as Profile likelihood confidence intervals and Wald confidence intervals correlations of parameter estimates covariances of parameter estimates observation statistics, such as influence diagnostics, predicted values and confidence intervals, and residuals multiple comparisons for classification effects exact tests, which are available only for binomial distributions with a logit link function or a Poisson distribution with a log link function.
Plots
You can select the plots to display in the output. If you choose to display multiple plots, you can display these plots individually or as a panel. Here are some plots that you can include in your results: predicted plots influence plots, such as Cook’s D by observation number and DFBETA by observation number plots of residuals, deviance residuals, standardized deviance residuals, Pearson residuals, standardized Pearson residuals, standardized Pearson residuals, and likelihood residuals.

Setting the Output Options

You can specify whether to create an output data set. You can also specify the values to include in the output data set. You can include predicted values, residuals, influence statistics, and the standard error of the linear predictor in the output data set.