Partager via

rxLogit: Logistic Regression

  • Article

Description

Use rxLogit to fit logistic regression models for small or large data.

Usage

  rxLogit(formula, data, pweights = NULL, fweights = NULL, cube = FALSE,
          cubePredictions = FALSE, variableSelection = list(), rowSelection = NULL, 
          transforms = NULL, transformObjects = NULL,
          transformFunc = NULL, transformVars = NULL, 
          transformPackages = NULL, transformEnvir = NULL,   
          dropFirst = FALSE, dropMain = rxGetOption("dropMain"),
          covCoef = FALSE, covData = FALSE, 
          initialValues = NULL, 
          coefLabelStyle = rxGetOption("coefLabelStyle"),
          blocksPerRead = rxGetOption("blocksPerRead"), 
          maxIterations = 25, coeffTolerance = 1e-06, 
          gradientTolerance = 1e-06, objectiveFunctionTolerance = 1e-08,
          reportProgress = rxGetOption("reportProgress"), verbose = 0,
          computeContext = rxGetOption("computeContext"),
          ...)

Arguments

formula

formula as described in rxFormula. Dependent variable must be binary. It can be a logical variable, a factor with only two categories, or a numeric variable with values in the range (0,1). In the latter case it will be converted to a logical.

data

either a data source object, a character string specifying a .xdf file, or a data frame object.

pweights

character string specifying the variable to use as probability weights for the observations.

fweights

character string specifying the variable to use as frequency weights for the observations.

cube

logical flag. If TRUE and the first term of the predictor variables is categorical (a factor or an interaction of factors), the regression is performed by applying the Frisch-Waugh-Lovell Theorem, which uses a partitioned inverse to save on computation time and memory. See Details section below.

cubePredictions

logical flag. If TRUE and cube is TRUE the estimated model is evaluated (predicted) for each cell in the cube, fixing the non-cube variables in the model at their mean values, and these predictions are included in the countDF component of the returned value. This may be time and memory intensive for large models. See Details section below.

variableSelection

a list specifying various parameters that control aspects of stepwise regression. If it is an empty list (default), no stepwise model selection will be performed. If not, stepwise regression will be performed and cube must be FALSE. See rxStepControl for details.

rowSelection

name of a logical variable in the data set (in quotes) or a logical expression using variables in the data set to specify row selection. For example, rowSelection = "old" will use only observations in which the value of the variable old is TRUE. rowSelection = (age > 20) & (age < 65) & (log(income) > 10) will use only observations in which the value of the age variable is between 20 and 65 and the value of the log of the income variable is greater than 10. The row selection is performed after processing any data transformations (see the arguments transforms or transformFunc). As with all expressions, rowSelection can be defined outside of the function call using the expression function.

transforms

an expression of the form list(name = expression, ...) representing the first round of variable transformations. As with all expressions, transforms (or rowSelection) can be defined outside of the function call using the expression function.

transformObjects

a named list containing objects that can be referenced by transforms, transformsFunc, and rowSelection.

transformFunc

variable transformation function. See rxTransform for details.

transformVars

character vector of input data set variables needed for the transformation function. See rxTransform for details.

transformPackages

character vector defining additional R packages (outside of those specified in rxGetOption("transformPackages")) to be made available and preloaded for use in variable transformation functions, e.g., those explicitly defined in RevoScaleR functions via their transforms and transformFunc arguments or those defined implicitly via their formula or rowSelection arguments. The transformPackages argument may also be NULL, indicating that no packages outside rxGetOption("transformPackages") will be preloaded.

transformEnvir

user-defined environment to serve as a parent to all environments developed internally and used for variable data transformation. If transformEnvir = NULL, a new "hash" environment with parent baseenv() is used instead.

dropFirst

logical flag. If FALSE, the last level is dropped in all sets of factor levels in a model. If that level has no observations (in any of the sets), or if the model as formed is otherwise determined to be singular, then an attempt is made to estimate the model by dropping the first level in all sets of factor levels. If TRUE, the starting position is to drop the first level. Note that for cube regressions, the first set of factors is excluded from these rules and the intercept is dropped.

dropMain

logical value. If TRUE, main-effect terms are dropped before their interactions.

covCoef

logical flag. If TRUE and if cube is FALSE, the variance-covariance matrix of the regression coefficients is returned. Use the rxCovCoef function to obtain these data.

covData

logical flag. If TRUE and if cube is FALSE and if constant term is included in the formula, then the variance-covariance matrix of the data is returned. Use the rxCovData function to obtain these data.

initialValues

Starting values for the Iteratively Reweighted Least Squares algorithm used to estimate the model coefficients. Supported values for this argument come in four forms:

  • NA - Initial values of the parameters are computed by a weighted least squares step in which, if there is no user-specified weight, the expected value (mu) of the dependent variable is set to 0.75 if the binary dependent variable is 1, and to 0.25 if it is 0, and to a correspondingly weighted value otherwise. This is equivalent to the default option for the glm function's logistic regression initial values, and can sometimes lead to convergence when the default NULL option does not, though it may result in more iterations in other cases. If convergence fails with the default NULL option, estimation is restarted with NA, and vice versa.
  • NULL - (Default) The initial values will be estimated based on a linear regression. This can speed up convergence significantly in many cases. If the model fails to converge using these values, estimation is automatically re-started using the NA option for the initial values.
  • Scalar - A numeric scalar that will be replicated once for each of the coefficients in the model to form the initial values. Zeros are often good starting values.
  • Vector - A numeric vector of length K, where K is the number of model coefficients, including the intercept if any and including any that might be dropped from sets of dummies. This argument is most useful when a set of estimated coefficients is available from a similar set of data. Note that K can be found by running the model on a small number of observations and obtaining the number of coefficients from the output, say z, via length(z$coefficients).

coefLabelStyle

character string specifying the coefficient label style. The default is "Revo". If "R", R-compatible labels are created.

blocksPerRead

number of blocks to read for each chunk of data read from the data source.

maxIterations

maximum number of iterations.

coeffTolerance

convergence tolerance for coefficients. If the maximum absolute change in the coefficients (step), divided by the maximum absolute coefficient value, is less than or equal to this tolerance at the end of an iteration, the estimation is considered to have converged. To disable this test, set this value to 0.

gradientTolerance

This argument is deprecated. Use objectiveFunctionTolerance and coeffTolerance to control convergence.

objectiveFunctionTolerance

convergence tolerance for the objective function. If the absolute relative change in the deviance (-2.0 times log likelihood) is less than or equal to this tolerance at the end of an iteration, the estimation is considered to have converged. To disable this test, set this value to 0.

reportProgress

integer value with options:

  • 0: no progress is reported.
  • 1: the number of processed rows is printed and updated.
  • 2: rows processed and timings are reported.
  • 3: rows processed and all timings are reported.

verbose

integer value. If 0, no verbose output is printed during calculations. Integer values from 1 to 4 provide increasing amounts of information are provided.

computeContext

a valid RxComputeContext. The RxSpark and RxHadoopMR compute contexts distribute the computation among the nodes specified by the compute context; for other compute contexts, the computation is distributed if possible on the local computer.

...

additional arguments to be passed directly to the Revolution Compute Engine.

Details

The special function F() can be used in formula to force a variable to be interpreted as a factor.

When cube is TRUE, the Frisch-Waugh-Lovell (FWL) Theorem is applied to the model. The FWL approach parameterizes the model to include one coefficient for each category (a single factor level or combination of factor levels) instead of using an intercept in the model with contrasts for each of the factor combinations. Additionally when cube is TRUE, the output contains a countDF element representing the counts for each category. If cubePredictions is also TRUE, predicted values using means of conditional independent continuous variables and weighted coefficients of conditional independent categorical variables are also computed and included in countDF. This may be time and memory intensive for large models. If there are no conditional independent variables (outside of the cube), the predicted values are equivalent to the coefficients and will be included in countDF whenever cube is TRUE.

Value

an rxLogit object containing the following elements:

coefficients

named vector of coefficients.

covCoef

variance-covariance matrix for the regression coefficient estimates.

covData

variance-covariance matrix for the explanatory variables in the regression model.

condition.number

estimated reciprocal condition number of final weighted cross-product (X'WX) matrix.

rank

numeric rank of the fitted linear model.

aliased

logical vector specifying whether columns were dropped or not due to collinearity.

coef.std.error

standard errors of the coefficients.

coef.t.value

coefficients divided by their standard errors.

coef.p.value

p-values for coef.t.values, using the normal distribution (Pr(>|z|))

total.squares

Y'Y of raw Y's

f.pvalue

the p-value resulting from an F-test on the fitted model.

df

degrees of freedom, a 3-vector (p, n-p, p*), the last being the number of non-aliased coefficients.

y.names

names for dependent variables.

partitions

when cube is TRUE, partitioned results will also be returned.

fstatistics

(for models including non-intercept terms) a 3-vector with the value of the F-statistic with its numerator and denominator degrees of freedom.

params

parameters sent to Microsoft R Services Compute Engine.

formula

the model formula

call

the matched call.

countDF

when cube is TRUE, a data frame containing category counts for the cube portion will also be returned. If cubePredictions is also TRUE, predicted values for each group in the cube are included.

nValidObs

number of valid observations.

nMissingObs

number of missing observations.

deviance

minus twice the maximized log-likelihood (up to a constant)

dispersion

for rxLogit, which estimates a generalized linear model with family binomial, the dispersion is always 1.

Note

Logistic regression is computed using the Iteratively Reweighted Least Squares (IRLS) algorithm, which is equivalent to full maximum likelihood.

Author(s)

Microsoft Corporation Microsoft Technical Support

References

Frisch, Ragnar; Waugh, Frederick V., Partial Time Regressions as Compared with Individual Trends, Econometrica, 1 (4) (Oct., 1933), pp. 387-401.

Lovell, M., 1963, Seasonal adjustment of economic time series, Journal of the American Statistical Association, 58, pp. 993-1010.

Lovell, M., 2008, A Simple Proof of the FWL (Frisch,Waugh,Lovell) Theorem, Journal of Economic Education.

See Also

rxLinMod, rxTransform.

Examples


 # Compare rxLogit and glm, which produce similar results when contr.SAS is used
 infertLogit <- rxLogit(case ~ age + parity + education + spontaneous + induced,
                        data = infert)
 infertLogit
 summary(infertLogit)

 infertGLM <- glm(case ~ age + parity + education + spontaneous + induced,
                  data = infert, family = binomial(), contrasts = list(education = contr.SAS)) 
 summary(infertGLM)

 # Instead of using the equivalent of contr.SAS, estimate the parameters
 # for the categorical levels without contrasting against an intercept term.
 infertCubeLogit <-
     rxLogit(case ~ education + age + parity + spontaneous + induced,
             data = infert, cube = TRUE)

 # Define an ageGroup variable through a variable transformation list. 
 # Partition the age groups between 1 and 100 by 20 years. 
 rxLogit(case ~ education + ageGroup + parity + spontaneous + induced,
         transforms=list(ageGroup = cut(age, seq(1, 100, 20))),
         data = infert, cube = TRUE)

 summary(infertCubeLogit)