GeneralizedLinearModel

Generalized linear regression model class

Description

GeneralizedLinearModel is a fitted generalized linear regression model. A generalized linear regression model is a special class of nonlinear models that describe a nonlinear relationship between a response and predictors. A generalized linear regression model has generalized characteristics of a linear regression model. The response variable follows a normal, binomial, Poisson, gamma, or inverse Gaussian distribution with parameters including the mean response μ. A link function f defines the relationship between μ and the linear combination of predictors.

Use the properties of a GeneralizedLinearModel object to investigate a fitted generalized linear regression model. The object properties include information about coefficient estimates, summary statistics, fitting method, and input data. Use the object functions to predict responses and to modify, evaluate, and visualize the model.

Creation

Create a GeneralizedLinearModel object by using fitglm or stepwiseglm.

fitglm fits a generalized linear regression model to data using a fixed model specification. Use addTerms, removeTerms, or step to add or remove terms from the model. Alternatively, use stepwiseglm to fit a model using stepwise generalized linear regression.

Properties

Coefficient Estimates

CoefficientCovariance — Covariance matrix of coefficient estimates

numeric matrix

This property is read-only.

Covariance matrix of coefficient estimates, specified as a p-by-p matrix of numeric values. p is the number of coefficients in the fitted model, as given by NumCoefficients.

For details, see Coefficient Standard Errors and Confidence Intervals.

Data Types: single | double

CoefficientNames — Coefficient names

cell array of character vectors

This property is read-only.

Coefficient names, specified as a cell array of character vectors, each containing the name of the corresponding term.

Data Types: cell

Coefficients — Coefficient values

table

This property is read-only.

Coefficient values, specified as a table. Coefficients contains one row for each coefficient and these columns:

• Estimate — Estimated coefficient value

• SE — Standard error of the estimate

• tStat — t-statistic for a two-sided test with the null hypothesis that the coefficient is zero

• pValue — p-value for the t-statistic

Use coefTest to perform linear hypothesis tests on the coefficients. Use coefCI to find the confidence intervals of the coefficient estimates.

To obtain any of these columns as a vector, index into the property using dot notation. For example, obtain the estimated coefficient vector in the model mdl:

beta = mdl.Coefficients.Estimate

Data Types: table

NumCoefficients — Number of model coefficients

positive integer

This property is read-only.

Number of model coefficients, specified as a positive integer. NumCoefficients includes coefficients that are set to zero when the model terms are rank deficient.

Data Types: double

NumEstimatedCoefficients — Number of estimated coefficients

positive integer

This property is read-only.

Number of estimated coefficients in the model, specified as a positive integer. NumEstimatedCoefficients does not include coefficients that are set to zero when the model terms are rank deficient. NumEstimatedCoefficients is the degrees of freedom for regression.

Data Types: double

Summary Statistics

Deviance — Deviance of fit

numeric value

This property is read-only.

Deviance of the fit, specified as a numeric value. The deviance is useful for comparing two models when one model is a special case of the other model. The difference between the deviance of the two models has a chi-square distribution with degrees of freedom equal to the difference in the number of estimated parameters between the two models. For more information, see Deviance.

Data Types: single | double

DFE — Degrees of freedom for error

positive integer

This property is read-only.

Degrees of freedom for the error (residuals), equal to the number of observations minus the number of estimated coefficients, specified as a positive integer.

Data Types: double

Diagnostics — Observation diagnostics

table

This property is read-only.

Observation diagnostics, specified as a table that contains one row for each observation and the columns described in this table.

Column	Meaning	Description
Leverage	Diagonal elements of HatMatrix	Leverage for each observation indicates to what extent the fit is determined by the observed predictor values. A value close to 1 indicates that the fit is largely determined by that observation, with little contribution from the other observations. A value close to 0 indicates that the fit is largely determined by the other observations. For a model with P coefficients and N observations, the average value of Leverage is P/N. A Leverage value greater than 2*P/N indicates high leverage.
CooksDistance	Cook's distance of scaled change in fitted values	CooksDistance is a measure of scaled change in fitted values. An observation with CooksDistance greater than three times the mean Cook's distance can be an outlier.
HatMatrix	Projection matrix to compute fitted from observed responses	HatMatrix is an N-by-N matrix such that Fitted = HatMatrix*Y, where Y is the response vector and Fitted is the vector of fitted response values.

The software computes these values on the scale of the linear combination of the predictors, stored in the LinearPredictor field of the Fitted and Residuals properties. For example, the software computes the diagnostic values by using the fitted response and adjusted response values from the model mdl.

Yfit = mdl.Fitted.LinearPredictor
Yadjusted = mdl.Fitted.LinearPredictor + mdl.Residuals.LinearPredictor

Diagnostics contains information that is helpful in finding outliers and influential observations. For more details, see Leverage, Cook’s Distance, and Hat Matrix.

Use plotDiagnostics to plot observation diagnostics.

Rows not used in the fit because of missing values (in ObservationInfo.Missing) or excluded values (in ObservationInfo.Excluded) contain NaN values in the CooksDistance column and zeros in the Leverage and HatMatrix columns.

To obtain any of these columns as an array, index into the property using dot notation. For example, obtain the hat matrix in the model mdl:

HatMatrix = mdl.Diagnostics.HatMatrix;

Data Types: table

Dispersion — Scale factor of variance of response

numeric scalar

This property is read-only.

Scale factor of the variance of the response, specified as a numeric scalar.

If the 'DispersionFlag' name-value pair argument of fitglm or stepwiseglm is true, then the function estimates the Dispersion scale factor in computing the variance of the response. The variance of the response equals the theoretical variance multiplied by the scale factor.

For example, the variance function for the binomial distribution is p(1–p)/n, where p is the probability parameter and n is the sample size parameter. If Dispersion is near 1, the variance of the data appears to agree with the theoretical variance of the binomial distribution. If Dispersion is larger than 1, the data set is “overdispersed” relative to the binomial distribution.

Data Types: double

DispersionEstimated — Flag to indicate use of dispersion scale factor

logical value

This property is read-only.

Flag to indicate whether fitglm used the Dispersion scale factor to compute standard errors for the coefficients in Coefficients.SE, specified as a logical value. If DispersionEstimated is false, fitglm used the theoretical value of the variance.

• DispersionEstimated can be false only for the binomial and Poisson distributions.

• Set DispersionEstimated by setting the 'DispersionFlag' name-value pair argument of fitglm or stepwiseglm.

Data Types: logical

Fitted — Fitted response values based on input data

table

This property is read-only.

Fitted (predicted) values based on the input data, specified as a table that contains one row for each observation and the columns described in this table.

Column	Description
Response	Predicted values on the scale of the response
LinearPredictor	Predicted values on the scale of the linear combination of the predictors (same as the link function applied to the Response fitted values)
Probability	Fitted probabilities (included only with the binomial distribution)

To obtain any of these columns as a vector, index into the property using dot notation. For example, obtain the vector f of fitted values on the response scale in the model mdl:

f = mdl.Fitted.Response

Use predict to compute predictions for other predictor values, or to compute confidence bounds on Fitted.

Data Types: table

LogLikelihood — Loglikelihood

numeric value

This property is read-only.

Loglikelihood of the model distribution at the response values, specified as a numeric value. The mean is fitted from the model, and other parameters are estimated as part of the model fit.

Data Types: single | double

ModelCriterion — Criterion for model comparison

structure

This property is read-only.

Criterion for model comparison, specified as a structure with these fields:

• AIC — Akaike information criterion. AIC = –2*logL + 2*m, where logL is the loglikelihood and m is the number of estimated parameters.

• AICc — Akaike information criterion corrected for the sample size. AICc = AIC + (2*m*(m + 1))/(n – m – 1), where n is the number of observations.

• BIC — Bayesian information criterion. BIC = –2*logL + m*log(n).

• CAIC — Consistent Akaike information criterion. CAIC = –2*logL + m*(log(n) + 1).

Information criteria are model selection tools that you can use to compare multiple models fit to the same data. These criteria are likelihood-based measures of model fit that include a penalty for complexity (specifically, the number of parameters). Different information criteria are distinguished by the form of the penalty.

When you compare multiple models, the model with the lowest information criterion value is the best-fitting model. The best-fitting model can vary depending on the criterion used for model comparison.

To obtain any of the criterion values as a scalar, index into the property using dot notation. For example, obtain the AIC value aic in the model mdl:

aic = mdl.ModelCriterion.AIC

Data Types: struct

Residuals — Residuals for fitted model

table

This property is read-only.

Residuals for the fitted model, specified as a table that contains one row for each observation and the columns described in this table.

Column	Description
Raw	Observed minus fitted values
LinearPredictor	Residuals on the linear predictor scale, equal to the adjusted response value minus the fitted linear combination of the predictors. For more information about the adjusted response, see Adjusted Response.
Pearson	Raw residuals divided by the estimated standard deviation of the response
Anscombe	Residuals defined on transformed data with the transformation selected to remove skewness
Deviance	Residuals based on the contribution of each observation to the deviance

Rows not used in the fit because of missing values (in ObservationInfo.Missing) contain NaN values.

To obtain any of these columns as a vector, index into the property using dot notation. For example, obtain the ordinary raw residual vector r in the model mdl:

r = mdl.Residuals.Raw

Data Types: table

Rsquared — R-squared value for model

structure

This property is read-only.

R-squared value for the model, specified as a structure with five fields.

Field	Description	Equation
Ordinary	Ordinary (unadjusted) R-squared	$$R_{\text{Ordinary}}^2=1-\frac{\text{SSE}}{\text{SST}}$$ SSE is the sum of squared errors, and SST is the total sum of squared deviations of the response vector from the mean of the response vector.
Adjusted	R-squared adjusted for the number of coefficients	$$R_{\text{Adjusted}}^2=1-\frac{\text{SSE}}{\text{SST}}\cdot \frac{N-1}{\text{DFE}}$$ N is the number of observations (NumObservations), and DFE is the degrees of freedom for the error (residuals).
LLR	Loglikelihood ratio	$$R_{\text{LLR}}^2=1-\frac{L}{L_0}$$ L is the loglikelihood of the fitted model (LogLikelihood), and L0 is the loglikelihood of a model that includes only a constant term. R2LLR is the McFadden pseudo R-squared value [1] for logistic regression models.
Deviance	Deviance R-squared	$$R_{\text{Deviance}}^2=1-\frac{D}{D_0}$$ D is the deviance of the fitted model (Deviance), and D0 is the deviance of a model that includes only a constant term.
AdjGeneralized	Adjusted generalized R-squared	$$R_{\text{AdjGeneralized}}^2=\frac{1-\exp \left(\frac{2\left(L_0-L\right)}{N}\right)}{1-\exp \left(\frac{2L_0}{N}\right)}$$ R2AdjGeneralized is the Nagelkerke adjustment [2] to a formula proposed by Maddala [3], Cox and Snell [4], and Magee [5] for logistic regression models.

To obtain any of these values as a scalar, index into the property using dot notation. For example, to obtain the adjusted R-squared value in the model mdl, enter:

r2 = mdl.Rsquared.Adjusted

Data Types: struct

SSE — Sum of squared errors

numeric value

This property is read-only.

Sum of squared errors (residuals), specified as a numeric value. If the model was trained with observation weights, the sum of squares in the SSE calculation is the weighted sum of squares.

Data Types: single | double

SSR — Regression sum of squares

numeric value

This property is read-only.

Regression sum of squares, specified as a numeric value. SSR is equal to the sum of the squared deviations between the fitted values and the mean of the response. If the model was trained with observation weights, the sum of squares in the SSR calculation is the weighted sum of squares.

Data Types: single | double

SST — Total sum of squares

numeric value

This property is read-only.

Total sum of squares, specified as a numeric value. SST is equal to the sum of squared deviations of the response vector y from the mean(y). If the model was trained with observation weights, the sum of squares in the SST calculation is the weighted sum of squares.

Data Types: single | double

Fitting Information

Steps — Stepwise fitting information

structure

This property is read-only.

Stepwise fitting information, specified as a structure with the fields described in this table.

Field	Description
Start	Formula representing the starting model
Lower	Formula representing the lower bound model. The terms in Lower must remain in the model.
Upper	Formula representing the upper bound model. The model cannot contain more terms than Upper.
Criterion	Criterion used for the stepwise algorithm, such as 'sse'
PEnter	Threshold for Criterion to add a term
PRemove	Threshold for Criterion to remove a term
History	Table representing the steps taken in the fit

The History table contains one row for each step, including the initial fit, and the columns described in this table.

Column	Description
Action	Action taken during the step: 'Start' — First step 'Add' — A term is added 'Remove' — A term is removed
TermName	If Action is 'Start', TermName specifies the starting model specification. If Action is 'Add' or 'Remove', TermName specifies the term added or removed in the step.
Terms	Model specification in a Terms Matrix
DF	Regression degrees of freedom after the step
delDF	Change in regression degrees of freedom from the previous step (negative for steps that remove a term)
Deviance	Deviance (residual sum of squares) at the step (only for a generalized linear regression model)
FStat	F-statistic that leads to the step
PValue	p-value of the F-statistic

The structure is empty unless you fit the model using stepwise regression.

Data Types: struct

Input Data

Distribution — Generalized distribution information

structure

This property is read-only.

Generalized distribution information, specified as a structure with the fields described in this table.

Field	Description
Name	Name of the distribution: 'normal', 'binomial', 'poisson', 'gamma', or 'inverse gaussian'
DevianceFunction	Function that computes the components of the deviance as a function of the fitted parameter values and the response values
VarianceFunction	Function that computes the theoretical variance for the distribution as a function of the fitted parameter values. When DispersionEstimated is true, the software multiplies the variance function by Dispersion in the computation of the coefficient standard errors.

Data Types: struct

Formula — Model information

LinearFormula object

This property is read-only.

Model information, specified as a LinearFormula object.

Display the formula of the fitted model mdl using dot notation:

mdl.Formula

LikelihoodPenalty — Penalty for likelihood estimate

"none" (default) | "jeffreys-prior"

Penalty for the likelihood estimate, specified as "none" or "jeffreys-prior".

• "none" — The likelihood estimate is not penalized during model fitting.

• "jeffreys-prior" — The likelihood estimate is penalized using the Jeffreys prior.

For logistic models, setting LikelihoodPenalty to "jeffreys-prior" is called Firth's regression. To reduce the coefficient estimate bias when you have a small number of samples, or when you are performing binomial (logistic) regression on a separable data set, set LikelihoodPenalty to "jeffreys-prior" during training.

Example: LikelihoodPenalty="jeffreys-prior"

Data Types: char | string

Link — Link function

structure

This property is read-only.

Link function, specified as a structure with the fields described in this table.

Field	Description
Name	Name of the link function, specified as a character vector. If you specify the link function using a function handle, then Name is ''.
Link	Function f that defines the link function, specified as a function handle
Derivative	Derivative of f, specified as a function handle
Inverse	Inverse of f, specified as a function handle

The link function is a function f that links the distribution parameter μ to the fitted linear combination Xb of the predictors:

f(μ) = Xb.

Data Types: struct

NumObservations — Number of observations

positive integer

This property is read-only.

Number of observations the fitting function used in fitting, specified as a positive integer. NumObservations is the number of observations supplied in the original table, dataset, or matrix, minus any excluded rows (set with the 'Exclude' name-value pair argument) or rows with missing values.

Data Types: double

NumPredictors — Number of predictor variables

positive integer

This property is read-only.

Number of predictor variables used to fit the model, specified as a positive integer.

Data Types: double

NumVariables — Number of variables

positive integer

This property is read-only.

Number of variables in the input data, specified as a positive integer. NumVariables is the number of variables in the original table or dataset, or the total number of columns in the predictor matrix and response vector.

NumVariables also includes any variables that are not used to fit the model as predictors or as the response.

Data Types: double

ObservationInfo — Observation information

table

This property is read-only.

Observation information, specified as an n-by-4 table, where n is equal to the number of rows of input data. ObservationInfo contains the columns described in this table.

Column	Description
Weights	Observation weights, specified as a numeric value. The default value is 1.
Excluded	Indicator of excluded observations, specified as a logical value. The value is true if you exclude the observation from the fit by using the 'Exclude' name-value pair argument.
Missing	Indicator of missing observations, specified as a logical value. The value is true if the observation is missing.
Subset	Indicator of whether or not the fitting function uses the observation, specified as a logical value. The value is true if the observation is not excluded or missing, meaning the fitting function uses the observation.

To obtain any of these columns as a vector, index into the property using dot notation. For example, obtain the weight vector w of the model mdl:

w = mdl.ObservationInfo.Weights

Data Types: table

ObservationNames — Observation names

cell array of character vectors

This property is read-only.

Observation names, specified as a cell array of character vectors containing the names of the observations used in the fit.

• If the fit is based on a table or dataset containing observation names, ObservationNames uses those names.

• Otherwise, ObservationNames is an empty cell array.

Data Types: cell

Offset — Offset variable

numeric vector

This property is read-only.

Offset variable, specified as a numeric vector with the same length as the number of rows in the data. Offset is passed from fitglm or stepwiseglm in the 'Offset' name-value pair argument. The fitting functions use Offset as an additional predictor variable with a coefficient value fixed at 1. In other words, the formula for fitting is

f(μ) ~ Offset + (terms involving real predictors)

where f is the link function. The Offset predictor has coefficient 1.

For example, consider a Poisson regression model. Suppose the number of counts is known for theoretical reasons to be proportional to a predictor A. By using the log link function and by specifying log(A) as an offset, you can force the model to satisfy this theoretical constraint.

Data Types: double

PredictorNames — Names of predictors used to fit model

cell array of character vectors

This property is read-only.

Names of predictors used to fit the model, specified as a cell array of character vectors.

Data Types: cell

ResponseName — Response variable name

character vector

This property is read-only.

Response variable name, specified as a character vector.

Data Types: char

VariableInfo — Information about variables

table

This property is read-only.

Information about variables contained in Variables, specified as a table with one row for each variable and the columns described in this table.

Column	Description
Class	Variable class, specified as a cell array of character vectors, such as 'double' and 'categorical'
Range	Variable range, specified as a cell array of vectors Continuous variable — Two-element vector [min,max], the minimum and maximum values Categorical variable — Vector of distinct variable values
InModel	Indicator of which variables are in the fitted model, specified as a logical vector. The value is true if the model includes the variable.
IsCategorical	Indicator of categorical variables, specified as a logical vector. The value is true if the variable is categorical.

VariableInfo also includes any variables that are not used to fit the model as predictors or as the response.

Data Types: table

VariableNames — Names of variables

cell array of character vectors

This property is read-only.

Names of variables, specified as a cell array of character vectors.

• If the fit is based on a table or dataset, this property provides the names of the variables in the table or dataset.

• If the fit is based on a predictor matrix and response vector, VariableNames contains the values specified by the 'VarNames' name-value pair argument of the fitting method. The default value of 'VarNames' is {'x1','x2',...,'xn','y'}.

VariableNames also includes any variables that are not used to fit the model as predictors or as the response.

Data Types: cell

Variables — Input data

table

This property is read-only.

Input data, specified as a table. Variables contains both predictor and response values. If the fit is based on a table or dataset array, Variables contains all the data from the table or dataset array. Otherwise, Variables is a table created from the input data matrix X and the response vector y.

Variables also includes any variables that are not used to fit the model as predictors or as the response.

Data Types: table

Object Functions

Create CompactGeneralizedLinearModel

compact

Compact generalized linear regression model

Add or Remove Terms from Generalized Linear Model

addTerms	Add terms to generalized linear regression model
removeTerms	Remove terms from generalized linear regression model
step	Improve generalized linear regression model by adding or removing terms

Predict Responses

feval	Predict responses of generalized linear regression model using one input for each predictor
predict	Predict responses of generalized linear regression model
random	Simulate responses with random noise for generalized linear regression model

Evaluate Generalized Linear Model

coefCI	Confidence intervals of coefficient estimates of generalized linear regression model
coefTest	Linear hypothesis test on generalized linear regression model coefficients
devianceTest	Analysis of deviance for generalized linear regression model
partialDependence	Compute partial dependence

Visualize Generalized Linear Model and Summary Statistics

plotDiagnostics	Plot observation diagnostics of generalized linear regression model
plotPartialDependence	Create partial dependence plot (PDP) and individual conditional expectation (ICE) plots
plotResiduals	Plot residuals of generalized linear regression model
plotSlice	Plot of slices through fitted generalized linear regression surface

Gather Properties of Generalized Linear Model

gather

Gather properties of Statistics and Machine Learning Toolbox object from GPU

Examples

Create Generalized Linear Regression Model

Fit a logistic regression model of the probability of smoking as a function of age, weight, and sex, using a two-way interaction model.

Load the hospital data set.

load hospital

Convert the dataset array to a table.

tbl = dataset2table(hospital);

Specify the model using a formula that includes two-way interactions and lower-order terms.

modelspec = 'Smoker ~ Age*Weight*Sex - Age:Weight:Sex';

Create the generalized linear model.

mdl = fitglm(tbl,modelspec,'Distribution','binomial')

mdl = 
Generalized linear regression model:
    logit(Smoker) ~ 1 + Sex*Age + Sex*Weight + Age*Weight
    Distribution = Binomial

Estimated Coefficients:
                        Estimate         SE         tStat      pValue 
                       ___________    _________    ________    _______

    (Intercept)            -6.0492       19.749     -0.3063    0.75938
    Sex_Male               -2.2859       12.424    -0.18399    0.85402
    Age                    0.11691      0.50977     0.22934    0.81861
    Weight                0.031109      0.15208     0.20455    0.83792
    Sex_Male:Age          0.020734      0.20681     0.10025    0.92014
    Sex_Male:Weight        0.01216     0.053168     0.22871     0.8191
    Age:Weight         -0.00071959    0.0038964    -0.18468    0.85348


100 observations, 93 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 5.07, p-value = 0.535

The large p-value indicates that the model might not differ statistically from a constant.

Create Generalized Linear Regression Model Using Stepwise Regression

Create response data using three of 20 predictor variables, and create a generalized linear model using stepwise regression from a constant model to see if stepwiseglm finds the correct predictors.

Generate sample data that has 20 predictor variables. Use three of the predictors to generate the Poisson response variable.

rng default % for reproducibility
X = randn(100,20);
mu = exp(X(:,[5 10 15])*[.4;.2;.3] + 1);
y = poissrnd(mu);

Fit a generalized linear regression model using the Poisson distribution. Specify the starting model as a model that contains only a constant (intercept) term. Also, specify a model with an intercept and linear term for each predictor as the largest model to consider as the fit by using the 'Upper' name-value pair argument.

mdl =  stepwiseglm(X,y,'constant','Upper','linear','Distribution','poisson')

1. Adding x5, Deviance = 134.439, Chi2Stat = 52.24814, PValue = 4.891229e-13
2. Adding x15, Deviance = 106.285, Chi2Stat = 28.15393, PValue = 1.1204e-07
3. Adding x10, Deviance = 95.0207, Chi2Stat = 11.2644, PValue = 0.000790094

mdl = 
Generalized linear regression model:
    log(y) ~ 1 + x5 + x10 + x15
    Distribution = Poisson

Estimated Coefficients:
                   Estimate       SE       tStat       pValue  
                   ________    ________    ______    __________

    (Intercept)     1.0115     0.064275    15.737    8.4217e-56
    x5             0.39508     0.066665    5.9263    3.0977e-09
    x10            0.18863      0.05534    3.4085     0.0006532
    x15            0.29295     0.053269    5.4995    3.8089e-08


100 observations, 96 error degrees of freedom
Dispersion: 1
Chi^2-statistic vs. constant model: 91.7, p-value = 9.61e-20

stepwiseglm finds the three correct predictors: x5, x10, and x15.

More About

Adjusted Response

The adjusted response for an observation is the first order Taylor expansion of the link function about a fitted response $${\widehat{y}}_i$$, evaluated at the observed response $$y_i$$. The adjusted response is given by

$$z_i=X_i\beta +offset+(y_i-{\widehat{y}}_i)g\text{'}({\widehat{y}}_i),$$

where

• $$z_i$$ is the adjusted response, corresponding to the ith observation.

• $$X_i$$ is the ith row in the matrix of predictor data.

• β is the column vector of model coefficients specified in mdl.Coefficients.

• offset is the offset of the model specified in mdl.Offset.

• $$y_i$$ is the observed response for the ith observation.

• $${\widehat{y}}_i$$ is the fitted response value for the ith observation.

• g is the link function in mdl.Link.

Canonical Link Function

The default link function for a generalized linear model is the canonical link function. You can specify a link function when you fit a model with fitglm or stepwiseglm by using the 'Link' name-value pair argument.

Distribution	Canonical Link Function Name	Link Function	Mean (Inverse) Function
'normal'	'identity'	f(μ) = μ	μ = Xb
'binomial'	'logit'	f(μ) = log(μ/(1 – μ))	μ = exp(Xb) / (1 + exp(Xb))
'poisson'	'log'	f(μ) = log(μ)	μ = exp(Xb)
'gamma'	-1	f(μ) = 1/μ	μ = 1/(Xb)
'inverse gaussian'	-2	f(μ) = 1/μ2	μ = (Xb)–1/2

Cook’s Distance

Cook’s distance is the scaled change in fitted values, which is useful for identifying outliers in the observations for predictor variables. Cook’s distance shows the influence of each observation on the fitted response values. An observation with Cook’s distance larger than three times the mean Cook’s distance might be an outlier.

The Cook’s distance D_i of observation i is

$$D_i=w_i\frac{e_i^2}{p\widehat{\phi }}\frac{h_{ii}}{{\left(1-h_{ii}\right)}^2},$$

where

• $$\widehat{\phi }$$ is the dispersion parameter (estimated or theoretical).

• e_i is the linear predictor residual, $$g\left(y_i\right)-x_i\widehat{\beta }$$, where

  • g is the link function.

  • y_i is the observed response.

  • x_i is the observation.

  • $$\widehat{\beta }$$ is the estimated coefficient vector.

• p is the number of coefficients in the regression model.

• h_ii is the ith diagonal element of the Hat Matrix H.

Leverage

Leverage is a measure of the effect of a particular observation on the regression predictions due to the position of that observation in the space of the inputs.

The leverage of observation i is the value of the ith diagonal term h_ii of the hat matrix H. Because the sum of the leverage values is p (the number of coefficients in the regression model), an observation i can be considered an outlier if its leverage substantially exceeds p/n, where n is the number of observations.

Hat Matrix

The hat matrix is a projection matrix that projects the vector of response observations onto the vector of predictions.

The hat matrix H is defined in terms of the data matrix X and a diagonal weight matrix W:

H = X(X^TWX)^–1X^TW^T.

W has diagonal elements w_i:

$$w_i=\frac{g^{\prime }\left({\mu }_i\right)}{\sqrt{V\left({\mu }_i\right)}},$$

where

• g is the link function mapping y_i to x_ib.

• $$g^{\prime }$$ is the derivative of the link function g.

• V is the variance function.

• μ_i is the ith mean.

The diagonal elements H_ii satisfy

$$\begin{array}{l}0\le h_{ii}\le 1\\ {\displaystyle \sum _{i=1}^n{h}_{ii}}=p,\end{array}$$

where n is the number of observations (rows of X), and p is the number of coefficients in the regression model.

Deviance

Deviance is a generalization of the residual sum of squares. It measures the goodness of fit compared to a saturated model.

The deviance of a model M₁ is twice the difference between the loglikelihood of the model M₁ and the saturated model M_s. A saturated model is a model with the maximum number of parameters that you can estimate.

For example, if you have n observations (y_i, i = 1, 2, ..., n) with potentially different values for X_i^Tβ, then you can define a saturated model with n parameters. Let L(b,y) denote the maximum value of the likelihood function for a model with the parameters b. Then the deviance of the model M₁ is

$$-2\left(\log L\left(b_1,y\right)-\log L\left(b_S,y\right)\right),$$

where b₁ and b_s contain the estimated parameters for the model M₁ and the saturated model, respectively. The deviance has a chi-squared distribution with n – p degrees of freedom, where n is the number of parameters in the saturated model and p is the number of parameters in the model M₁.

Assume you have two different generalized linear regression models M₁ and M₂, and M₁ has a subset of the terms in M₂. You can assess the fit of the models by comparing their deviances D₁ and D₂. The difference of the deviances is

$$\begin{array}{l}D=D_2-D_1=-2\left(\log L\left(b_2,y\right)-\log L\left(b_S,y\right)\right)+2\left(\log L\left(b_1,y\right)-\log L\left(b_S,y\right)\right)\\ =-2\left(\log L\left(b_2,y\right)-\log L\left(b_1,y\right)\right).\end{array}$$

Asymptotically, the difference D has a chi-squared distribution with degrees of freedom v equal to the difference in the number of parameters estimated in M₁ and M₂. You can obtain the p-value for this test by using 1  —  chi2cdf(D,v).

Typically, you examine D using a model M₂ with a constant term and no predictors. Therefore, D has a chi-squared distribution with p – 1 degrees of freedom. If the dispersion is estimated, the difference divided by the estimated dispersion has an F distribution with p – 1 numerator degrees of freedom and n – p denominator degrees of freedom.

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.

References

[1] McFadden, Daniel. "Conditional logit analysis of qualitative choice behavior." in Frontiers in Econometrics, edited by P. Zarembka,105–42. New York: Academic Press, 1974.

[2] Nagelkerke, N. J. D. "A Note on a General Definition of the Coefficient of Determination." Biometrika 78, no. 3 (1991): 691–92.

[3] Maddala, Gangadharrao S. Limited-Dependent and Qualitative Variables in Econometrics. Econometric Society Monographs. New York, NY: Cambridge University Press, 1983.

[4] Cox, D. R., and E. J. Snell. Analysis of Binary Data. 2nd ed. Monographs on Statistics and Applied Probability 32. London; New York: Chapman and Hall, 1989.

[5] Magee, Lonnie. "R 2 Measures Based on Wald and Likelihood Ratio Joint Significance Tests." The American Statistician 44, no. 3 (August 1990): 250–53.

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using MATLAB® Coder™.

Usage notes and limitations:

• The predict and random functions support code generation.

• When you fit a model by using fitglm or stepwiseglm, you cannot specify Link, Derivative, and Inverse fields of the 'Link' name-value pair argument as anonymous functions. That is, you cannot generate code using a generalized linear model that was created using anonymous functions for links. Instead, define functions for link components.

For more information, see Introduction to Code Generation.

GPU Arrays

Accelerate code by running on a graphics processing unit (GPU) using Parallel Computing Toolbox™.

Usage notes and limitations:

• The object functions of the GeneralizedLinearModel model fully support GPU arrays.

For more information, see Run MATLAB Functions on a GPU (Parallel Computing Toolbox).

Version History

Introduced in R2012a